Unsloth v0.1.910-beta

Unsloth is an open-source AI platform for running and training open models locally. It combines Unsloth Studio, a graphical interface for running and training models, with Unsloth Core, a code-based library for fine-tuning and machine learning workflows. It supports language, vision, audio, embedding, and diffusion models across Windows, Linux, WSL, and macOS.

The project is particularly focused on making model fine-tuning more efficient by reducing memory usage and improving training performance. It can also run local models and connect them to coding agents and other tools.

Features

Unsloth provides a graphical environment for downloading, running, and training open models. Users can search for models, run them locally, fine-tune them, and export the resulting models to formats such as GGUF and safetensors.

Key features include:

• Local LLM inference

• LLM fine-tuning

• Vision and multimodal models

• Audio models

• Embedding models

• Diffusion models

• GGUF support

• MLX support

• LoRA and other fine-tuning workflows

• Model export

• RAG

• Web search

• Tool calling

• Code execution

• MCP support

• Multi-GPU support

• Claude Code integration

• OpenAI Codex integration

• OpenCode integration

• Docker support

• Self-hosted web interface

• Python API and notebooks

Unsloth Start can connect local models to supported AI coding agents, allowing tools such as Claude Code and Codex to use models running through Unsloth.

The platform also supports several hardware backends. The project lists NVIDIA, AMD, Intel, CPU, and Vulkan support, although capabilities differ depending on the hardware and workload.

Download Unsloth v0.1.910-beta - Software Mirrors
Unsloth v0.1.910-beta for Windows Unsloth-Desktop-Windows.exe | 20.87 MB Unsloth-Desktop-Windows-ARM64.exe | 20.32 MB
Unsloth v0.1.910-beta for macOS Unsloth-Desktop-MacOS.dmg | 22.64 MB
Unsloth v0.1.910-beta for Linux Unsloth-Desktop-Ubuntu.deb | 24.61 MB Unsloth-Desktop-Ubuntu-ARM64.deb | 24.92 MB Unsloth-Desktop-Linux.AppImage | 171.28 MB
Others Download related to Unsloth v0.1.910-beta Unsloth-Desktop-ARM64.app.tar.gz | 22.47 MB
Unsloth v0.1.910-beta Source Code Unsloth v0.1.910-beta Source code (zip) Unsloth v0.1.910-beta Source code (tar.gz)
Unsloth v0.1.910-beta Release Notes: Recursive's Flint, MOSS-Image and NVIDIA's Nemotron Bolt 4B are here! Run them locally in Unsloth. Also new: a chat-only Unsloth Desktop install, MLX MoE models bigger than your Mac's memory and an OpenAI-compatible fine-tuning API. Highlights Bigger MoE models on Mac: MLX MoE models too big for memory keep experts on SSD, so Qwen3.8-Flash-Next runs on a 32 GB Mac. Chat-only install: a much smaller Unsloth Desktop for GGUF chat; add training later in Settings. Flint: Recursive's 40B coding-agent MoE, runs in 24 GB at 3-bit Guide MOSS-Image: 5.1B image model designed for top-tier generation GGUF + FP8 NVIDIA Nemotron 3.5 Bolt E4B: NVIDIA's new Gemma 4 E4B fine-tune, 63.2 on Tau2-Bench vs the original's 38.7 Guide MoE expert pruning: Export a smaller MoE model by removing 25% or 50% of its least useful experts with REAP. UI overhaul: More consistent UI across Unsloth Images + video Generate images containing English or Chinese text with OpenMOSS's 5.1B MOSS-Image, SFT or Base. GGUF · Base GGUF MOSS-Image, the new image model, runs on 16 GB GPUs; its GGUFs also run on Macs and computers without a supported GPU. Most image and video models too big for one GPU now split across your NVIDIA GPUs, if they fit, instead of streaming from system memory. The new Hires fix upscales and refines a render in one step on SDXL, FLUX.1, Qwen-Image and Z-Image. The new Compare grid renders one prompt while varying up to three settings, such as seed, CFG or LoRA weight, in labelled grids. SDXL safetensors get AUTOMATIC1111-style prompt weighting, CLIP skip, CFG rescale, variation seeds and textual inversion. Pasted AUTOMATIC1111, Forge or Civitai parameter blocks fill in the Images page settings. If SAM 3 isn't already downloaded, the Mask from text tool now uses SAM 3.1: smaller, with no Hugging Face login. Chat + Desktop Pick Chat only (GGUF) when first launching Unsloth Desktop for a much smaller install; add training later in Settings. Settings > Chat now lets you view, edit, clear or turn off the memory notes Unsloth sends to every chat. In document-search chats, Show passages lists the exact text the model got and opens each passage at its page. The Audio page can speak and transcribe via a saved API connection, such as OpenAI or a local server. Unsloth installs, updates and pip install -U unsloth now reject eMudhra's compromised emSign certificates, which let attackers impersonate websites. The optional vLLM and SGLang engines still trust them. Flint Run Recursive's Flint, a 40B MoE model with 2.5B active parameters for coding agents, with reasoning, tool use and 256K context. Guide · GGUF At 3-bit, Flint fits in 24 GB at short contexts. On SWE-bench Pro v2, Flint scores 89.7 vs Qwen 3.6's 73.8. NVIDIA Nemotron 3.5 Bolt E4B Run and fine-tune NVIDIA's Gemma Nemotron 4B, a Gemma 4 E4B reasoning tune for tool calling and instruction following. Gemma Nemotron 4B beats Gemma 4 E4B on 16 of 21 NVIDIA tests, gaining most in tool use and multi-turn chat. DPO on Gemma 4 E2B and E4B, including Gemma Nemotron 4B, no longer crashes. Training Fine-tuning Whisper, Canary, Moonshine, Dia and CTC speech models with FastModel no longer fails. Parakeet, Canary and CTC speech training uses up to 55% less memory. The new Books & Documents to Text recipe in Data Recipes turns ebooks and documents into a pretraining text dataset, no model needed. On Windows and Linux, the Export page can now prune 25% or 50% of a MoE model's least useful experts with REAP, at some quality cost. In Python, reap_prune runs REAP expert pruning up to 54x faster than llm-compressor's. unsloth train now trains vision and, except on Mac, audio models. Unsloth now supports OpenAI's fine-tuning API, so OpenAI's SDK can upload datasets and run supervised fine-tuning of local models. Performance MLX MoE models too big for your Mac keep experts on SSD, with slower replies. Qwen3.8-Flash-Next fits a 32 GB Mac; DeepSeek-V4-Flash and GLM-5.3-Flash need far less memory than their size. DPO and other training that pads sequences runs up to 3.8x faster with less memory on NVIDIA Ampere or newer. Qwen3-MoE, Gemma 3 and Gemma 4 training on GPUs without bf16 now uses up to 7x less attention memory with torch 2.13 or older. Coding agents + API In chat, models can now read text files and documents on connected MCP servers, not just use their tools. When launching coding agents, unsloth start no longer sends its Unsloth API key in plain text through your HTTP proxy to a local Unsloth. On Windows, unsloth start no longer breaks Claude Code hooks or Codex sign-in when launching them. Requests with an Unsloth API key can no longer run Python and terminal tools unsandboxed; only the Unsloth UI can allow that. The new Save for this model button in Chat settings makes your sampling settings that model's default for new chats and API requests. Models + hardware A BF16 gpt-oss checkpoint loaded with load_in_4bit = True no longer gets random expert and router weights. The model picker again lets you run models on Ryzen AI 300-series and newer NPUs, with no environment variable. On some Linux AMD PCs with a system-wide ROCm, GGUFs that ran on the CPU now run on the GPU again. Download Unsloth Desktop Unsloth Desktop is free and open source. Download it for: -- Platform -- Link -- Windows -- Download -- macOS -- Download -- Linux x64 / Ubuntu (deb) -- Download -- Linux ARM64 / Ubuntu 24.04+ (deb) -- Download -- Linux x64 (AppImage) -- Download -- Windows ARM64 -- Download What's Changed Bump install.sh / install.ps1 pins to unsloth>=2026.10.7, unsloth-zoo>=2026.10.7 by @danielhanchen in #14276 Studio: accept a quoted local training base and name single-file checkpoints in the refusal by @danielhanchen in #14236 QLoRA quant-state fallback: gate per weight, not a process-global flag by @danielhanchen in #14231 GatedDeltaNet fla op engages in training; large-head self-check survives a non-reentrant checkpoint by @danielhanchen in #14233 gpt-oss MXFP4 without triton_kernels on transformers 4.x: keep dense expert biases in the model dtype by @danielhanchen in #14275 FSDP2: gather for push_to_hub, refuse GGUF / torchao exports, size offloaded state and 4-bit counts per shard by @danielhanchen in #14205 Large head attention follow-ups: compiled kill switch, 4-bit shard self check, forced float32 skip by @danielhanchen in #14188 Optimizer parking outside inference mode; refuse fp16 params with unquantized adamw_flash moments by @danielhanchen in #14201 Review follow-ups: fused LoRA dropout/bias under compile and 4-bit, Q-GaLore offload, GRPO truncated support, CP save on transformers 4.x by @danielhanchen in #14226 CI: raise the Core zoo job cap from 25 to 35 minutes by @danielhanchen in #14240 Studio: change Parallel Slots from the API page's Slots readout by @danielhanchen in #14225 FSDP2 QLoRA: keep packed 4-bit bytes out of the loader dtype pass, fix merged_4bit save by @danielhanchen in #14165 GRPO: reduce the IS ratio metric per sequence for geometric modes and the env override by @danielhanchen in #14192 Retry the audio.cpp server on a new port when its port is taken by @danielhanchen in #14294 fix(rag): avoid long CPU stalls when archiving tool calls by @ATAMAH in #14098 Avoid unnecessary attachment reindexing on chat generation by @Imagineer99 in #14217 Studio: update external llama.cpp default port to 9931 by @oobabooga in #14266 Import fix: torchao 0.17+ on torch 2.6 (torch.utils._pytree.register_constant) by @danielhanchen in #14299 whisper.cpp: skip the macOS release walk when llama.cpp is ahead of every release by @danielhanchen in #14302 Studio: Library chat cards keep their column width and equal padding by @shimmyshimmer in #14305 Baseline the unsloth-zoo 2026.10.7 scan findings after review by @danielhanchen in #14311 Studio: SAM 3.1 for Mask from text, with pre-quantized INT8 / FP8 files by @danielhanchen in #14310 Studio: serve /v1/completions from managed vLLM and SGLang engines by @danielhanchen in #14227 Studio: logprobs on /v1/completions and /tokenize for MLX models by @danielhanchen in #14229 Studio: fix FLUX.3 workflows, controls, LoRAs and download progress by @oobabooga in #14312 Add a vLLM backend to unsloth eval by @danielhanchen in #14228 Model catalog network check: judge FLUX.3 Klein by the mirror the backend fetches by @danielhanchen in #14316 SAM 3 skeleton test: compare against a float32 load and find the RoPE tables by any spelling by @danielhanchen in #14318 Raise the startup bundle budget for the FLUX.3 controls (#14312) by @danielhanchen in #14319 Studio: open the Unsloth docs and Next.js sites in the browser panel by @NilayYadav in #14279 Studio: load ComfyUI fp8mixed text encoders (nvfp4 / mxfp8 layers) by @danielhanchen in #14321 Studio: fix Claude API reference pages in the browser panel by @wasimysaid in #14323 Studio: let the browser agent tick radios and checkboxes by their label by @NilayYadav in #14282 Native FSDP2: fully_shard per block and root for every model, LoRA, QLoRA and full finetuning (UNSLOTH_FSDP2=1) by @danielhanchen in #14273 Studio: suggest Unsloth Docs in the web UI browser panel by @shimmyshimmer in #14324 Studio Hub: rank Latest by trending against creation date, not last modified by @shimmyshimmer in #14329 Studio: modern toggles, 34 x 20 with a flat centred thumb by @shimmyshimmer in #14322 Raise the startup bundle budget for the modern toggles (#14322) by @danielhanchen in #14331 Point Unsloth docs links at unsloth.ai/docs instead of docs.unsloth.ai by @shimmyshimmer in #14333 Distributed tests: pick MASTER_PORT outside the ephemeral range, write vision results atomically by @danielhanchen in #14334 Studio: keep project sources when a knowledge base is picked by @danielhanchen in #14255 Desktop: offer a chat-only (GGUF) install and keep it through repairs by @danielhanchen in #14245 Studio: let the model read MCP server resources by @danielhanchen in #14243 Move the Unsloth Studio MLX pins to mlx-vlm 0.7.9 by @Lyxot in #14307 Prebuilt wheels: build causal-conv1d and mamba-ssm for Linux torch 2.10 to 2.12 by @danielhanchen in #14336 PEFT MoE conversion snapshot: add the Qwen3-Omni MoE text type and thinker aliases by @danielhanchen in #14344 OpenAPI loop test: build the real schema outside the measured window by @danielhanchen in #14346 Studio diffusion: multi-GPU component placement and CFG-parallel by @danielhanchen in #13725 Studio: honor UNSLOTH_HOME in installers by @oobabooga in #14293 Studio: make copy buttons on web pages work in the browser panel by @NilayYadav in #14284 fix(studio): MLX memory estimate no longer allocates the model, and sizes DeepSeek V4 / V4.1 by @Lyxot in #14325 Studio: keep runtime flash-attn installs wheel-only by @oobabooga in #14287 Studio: keep text typed into a browser panel page when switching tabs by @NilayYadav in #14280 Normalize role names when converting ChatML to Alpaca by @ShryukGrandhi in #14304 fix(studio): align Whisper training labels with decoder prefix by @taking-lying-flat in #14244 Studio: skip causal-conv1d for Qwen3.5 / 3.6 / Qwen3-Next when unsloth-zoo covers decode by @danielhanchen in #14148 Studio: warn before downloading a program sent without an extension by @NilayYadav in #14278 unsloth start (Windows): pass the sub-agent task over stdin so non-npm .cmd shims don't mangle it by @LeoBorcherding in #14300 Studio: open each Settings tab at the top instead of the previous tab's scroll offset by @danielhanchen in #14353 Studio: keep the page open when saving a bookmark by @NilayYadav in #14286 Studio: download prebuilt kernel wheels once by @oobabooga in #14051 Studio: honor width, height, steps, guidance, seed and negative_prompt on /v1/images/generations by @danielhanchen in #14351 Studio: API keys run tools in the sandbox; Full access needs the Unsloth UI by @danielhanchen in #14345 fix: preserve padding-free DPO sample boundaries in fast Llama/Qwen2 by @taking-lying-flat in #14251 Unsloth Studio: a JSONL reply that quotes no longer stalls the tool parser and fails the turn with RecursionError by @LeoBorcherding in #14097 Studio diffusion: pin Triton libdevice before the first compile so restarts hit the FX graph cache by @danielhanchen in #13713 Studio: stop recipe dialogs opening a hint tooltip on open by @keyiyu0331 in #13668 Studio: size sd.cpp's thread pool from the Apple fast cores by @danielhanchen in #13349 Studio: prune dead startup records and preserve process identity by @oobabooga in #14301 Studio diffusion: single-card CFG batching for a DiT streamed from host memory by @danielhanchen in #13787 Studio: Mamba-2 hybrids use the extension-free mamba-ssm wheel when unsloth-zoo runs their conv on Triton by @danielhanchen in #14160 Keep SFT kwargs passed beside a plain TrainingArguments instead of falling back to trl's defaults by @LeoBorcherding in #14094 Unsloth Studio (AMD/Windows): show an iGPU's own memory on Settings > About, not the system RAM it can borrow by @LeoBorcherding in #14317 Studio compile cache: keep every graph across re-saves, and save under torch's compile lock by @danielhanchen in #14341 fix: preserve sentence encoder adapters and export dense 16bit weights by @taking-lying-flat in #14246 Studio diffusion: cache runtime-quantized denoisers as local pre-quant checkpoints by @danielhanchen in #13715 load_in_fp8=True quantizes a model with no FP8 repo instead of silently loading it in bf16 by @LeoBorcherding in #14093 check_tokenizer: repair out-of-vocab added tokens on transformers 5.x, and raise when one can't be repaired by @LeoBorcherding in #14096 save_pretrained_gguf: stop rewriting a reused local checkpoint's tokenizer with a BOS-stripped template by @LeoBorcherding in #14095 Studio sd.cpp: explicit VRAM budget for MiniMax-H3 and a CUDA capability fallback by @danielhanchen in #13714 MoE grouped GEMM: require sm90 on every visible card for TMA and launch on the operands' device by @danielhanchen in #14358 Studio diffusion: run FLUX.3 base conditioning blocks once per render by @danielhanchen in #14367 Fix chat context counts and token usage details by @Sneakr in #14076 Installers: do not install generic aarch64 CUDA torch on Jetson before CUDA 13 by @danielhanchen in #14363 Stop fix_executorch from hiding an installed torchtune by @danielhanchen in #14375 Tests: keep gloo FSDP2 test ranks on the CPU on GPU hosts by @danielhanchen in #14383 fix: release the device cache through the active accelerator on the vLLM resume/teardown paths by @li-lizhe in #13438 Fix phone-deployment QAT for Gemma3 and LoRA, and write the matching ExecuTorch export config by @danielhanchen in #14376 Studio: run Ada FP8 precise accumulate through a unit-scale GEMM plus a fused scale epilogue by @danielhanchen in #13793 Studio: pin the FLUX.3 prompt upsampler to the image model's card under HIP masks and on Vulkan builds by @danielhanchen in #14356 NF4 decode on T4: use the bitsandbytes 4-bit GEMM above the GEMV row cap by @danielhanchen in #14386 Studio diffusion: record a prompt length's CUDA graph on the first render only by @danielhanchen in #14387 Studio: Load anyway for the unified-memory diffusion refusal, and a capacity fix for its pre-download check by @danielhanchen in #14379 Studio/CLI: report Rosetta, skip audio.cpp CUDA on pre-Turing GPUs, pass doctor on MPS/XPU, hide XPU from token preview by @danielhanchen in #14374 llama.cpp prebuilt: accept upstream win-rocm bundles, run native arm64 under Rosetta, source-build on Jetson by @danielhanchen in #14373 Gate XPU bf16 on native support, hint CUDA_DEVICE_ORDER on mixed NVIDIA hosts, read the Metal GPU family for MLX bf16 by @danielhanchen in #14371 Raise the startup bundle budget for the Rosetta chat-only messages (#14374) by @danielhanchen in #14407 Block compile: keep the stock forward under DeepSpeed ZeRO-3 by @danielhanchen in #14396 Studio: prefer discrete GPUs over unified-memory APUs in automatic GGUF placement; clamp Vulkan free to total by @danielhanchen in #14372 Refuse load_in_8bit with fast_inference instead of silently loading vLLM in 16-bit by @danielhanchen in #13019 Studio: retry json_object as json_schema when a custom provider (LM Studio) refuses it by @cedric66 in #11813 Studio: fix First Block Cache on FLUX.2 (dual-stream head, single-stream tail) by @danielhanchen in #14401 Studio: offload-first retry for an OOM VAE decode, fuse GroupNorm+SiLU on uncovered VAEs by structure by @danielhanchen in #13802 FP8 block dequant below sm89: decode e4m3fn bytes in one Triton pass by @danielhanchen in #14408 int4 packed: trace the training dequantize inside torch.compile by @danielhanchen in #14385 Block compile: find decoder layers structurally and wrap mixed stacks per class by @danielhanchen in #14397 Native FSDP2: keep blocks that share parameters in the root unit (Zamba2) by @danielhanchen in #14384 Studio: pick ROCm flash attention by a whole-block timing probe instead of a model table by @danielhanchen in #13792 Studio: encode Wan clips while the VAE decodes by @danielhanchen in #14405 FastLanguageModel LoRA / QLoRA: compile each decoder block as one region by @danielhanchen in #14398 Studio: run MiniMax-H3 PDD checkpoints with their 32-head output bank, refuse the ones that cannot run by @danielhanchen in #14410 DeepSpeed ZeRO-2 / ZeRO-3: stage detection at load, MoE and fused LoRA leaves, precision resync, reentrant checkpointing by @danielhanchen in #14394 Fuse the decode residual add into RMSNorm (one Triton launch per norm) by @danielhanchen in #14382 Tests: read checked-in sources as UTF-8 in three new test files by @danielhanchen in #14414 QAT test: spoof CUDA in the unpickling child on GPU-less runners by @danielhanchen in #14415 Studio: show the retrieved passages behind a document search by @yejonz in #13987 Studio video stream encode: a failed final flush falls back instead of hanging the render by @danielhanchen in #14413 GRPO under DeepSpeed ZeRO-3: score with the gathered head, no_grad instead of inference_mode by @danielhanchen in #14395 FP8 byte dequant: pass enable_fp_fusion at the launch by @danielhanchen in #14416 Studio: decode the Rosetta sysctl probe as UTF-8 by @danielhanchen in #14419 DDP: gradient_as_bucket_view for full finetuning, opt-in bf16 / fp16 gradient all-reduce by @danielhanchen in #14399 Studio no-torch runtime: pin certifi 2026.10.19, the floor fastmcp-slim now requires by @danielhanchen in #14418 Tests: run the precision template in one namespace, keep the block-compile packing test off leaked env by @danielhanchen in #14425 RoPE: multi-head tiles, Q and K in one launch, no clones (2-2.7x per call) by @danielhanchen in #14417 gpt-oss: load_in_4bit from a BF16 checkpoint no longer loads random experts and router by @danielhanchen in #14400 Studio Hub: list every image and video model under Recommended by @shimmyshimmer in #14421 Studio diffusion: VAE OOM under a component split, expanded timesteps in the CFG routers, and test fixes by @danielhanchen in #14420 Studio diffusion: match family keywords only on a local model's own path, and undo the torch 2.13 slice patch in the speed tests by @danielhanchen in #14435 Studio: split one diffusion denoiser across GPUs by block ranges by @danielhanchen in #14412 Installers (ROCm on WSL): no empty LD_LIBRARY_PATH entry in the profile drop-in by @danielhanchen in #14433 Studio: retire the seeded password on Desktop and keep API callers on Sandbox High by @danielhanchen in #14422 Studio: harden Data Recipe engine inputs by @danielhanchen in #14424 Studio: harden the server against stalled bodies, long login names, readable logs and managed-account fetches by @danielhanchen in #14426 Studio: managed dataset media decodes stay on readable files, and Arrow data files are validated by @danielhanchen in #14430 Studio: managed accounts cannot use another account's loaded model or export worker by @danielhanchen in #14432 ROCm: eager batch-1 NF4 decode takes the Triton GEMV (1.8x decode on gfx1151) by @danielhanchen in #14440 Dependabot: hold cairo-rs and windows-core with the GTK3 / WebView2 pins by @danielhanchen in #14441 Bump the npm-oxc-validator group in /studio/backend/core/data_recipe/oxc-validator with 3 updates by @dependabot[bot] in #13832 Bump the actions group across 1 directory with 8 updates by @dependabot[bot] in #13835 Studio: approval waits free their slot, Deep Research honours --disable-tools, first-run shutdown survives a read error by @danielhanchen in #14434 Studio: tighten managed-account tool confinement by @danielhanchen in #14423 Installers (ROCm on WSL): fetch the helper from the commit with the LD_LIBRARY_PATH fix by @danielhanchen in #14449 Studio: Linux torch 2.10 to 2.12 take our causal-conv1d and mamba-ssm wheels by @danielhanchen in #14338 Studio diffusion: hand compiled blocks the recorded tensor kind on graphless steps (no recompile on a new prompt length) by @danielhanchen in #14448 Studio Train: clearer selected tabs, toggles and target modules in dark mode by @shimmyshimmer in #14451 Studio: harden the tool sandbox around host launches, probes and workdir paths by @danielhanchen in #14429 Studio: show the Auto download reason only when it falls back to HTTPS by @shimmyshimmer in #14454 Studio: keep diagrams, redirect links and approval cards honest by @danielhanchen in #14427 Studio: Projects new-chat button, Projects home in the sidebar menu, even menu padding by @shimmyshimmer in #14456 Honour every_n_layers and drop redundant offload in gradient_checkpointing_enable by @danielhanchen in #14403 Gradient checkpointing: checkpoint_skip_layers = "auto" and UNSLOTH_GC_KEEP_LAYERS by @danielhanchen in #14404 Bump the npm-oxc-validator group in /studio/backend/core/data_recipe/oxc-validator with 3 updates by @dependabot[bot] in #14467 Studio: managed-account paths are checked as they are opened by @danielhanchen in #14428 Studio and Core: close remote-code gate gaps in Sentence Transformers, GGUF export, scans and tokenizer paths by @danielhanchen in #14437 Call the app Unsloth or Unsloth Studio, never a bare Studio by @shimmyshimmer in #14473 unsloth start dsh: warn when the loaded context leaves replies no room by @danielhanchen in #14352 Studio: truncate over-length /v1/embeddings inputs instead of failing by @danielhanchen in #14364 Studio: managed accounts keep the configured Decision checkpoint; signed-out Hub relay hides upstream errors by @danielhanchen in #14462 Studio: one slider look everywhere, filled with the accent by @shimmyshimmer in #14471 Unsloth Studio: redesigned API page, multi-model support and a Run settings fix by @shimmyshimmer in #14455 Studio: let browser pages open target=_blank links and show alerts by @danielhanchen in #14360 Bump the cargo-tauri group in /studio/src-tauri with 26 updates by @dependabot[bot] in #14468 Studio Images: say why LoRAs vanish after Speed=Auto compiles, offer a reload by @danielhanchen in #14359 Studio: train the system column of an Alpaca dataset by @NilayYadav in #14392 Studio: point Desktop Chat only users at Repair installation on the Export page by @oobabooga in #14438 Studio: count reclaimable active file cache as available host RAM under a cgroup limit by @danielhanchen in #14475 Studio: Z-Image unload test works on diffusers without ROPE_PER_DEVICE by @danielhanchen in #14478 Studio: keep Insert > Symbol characters when reading Word files by @NilayYadav in #14393 Core: per-user fp8 cache, SyntheticDataKit binds vLLM to loopback by @danielhanchen in #14460 Desktop and Windows installer: browser-panel capture prompt, private ACL for a custom Studio home by @danielhanchen in #14464 Studio: keep the chat model loaded when an API request loads a speech model by @danielhanchen in #14369 Studio: initialise the CUDA driver off the GIL so load-progress polls stay responsive by @danielhanchen in #14452 Fix: [Bug] Installer hard-codes PyPI.org instead of respecting PIP_INDEX_URL / UV_INDEX_URL by @SIDDARTHAREDDY8 in #13748 Dependabot: leave hand-kept pins to the version matrix and hold the GTK3 and deep-link crates by @danielhanchen in #14470 Studio: honor completion-only training for audio VLMs by @taking-lying-flat in #14402 Studio: never walk init's descendants when stopping a server by @danielhanchen in #14486 Studio: save files from download buttons inside browser panel pages by @NilayYadav in #14285 Studio: keep tools working after opening a chat file in its app by @NilayYadav in #14390 Studio: stop two Studio windows undoing each other's browser data by @NilayYadav in #14277 Sample Qwen3 models for the turn's thinking mode in unsloth chat by @NilayYadav in #14388 MoE grouped GEMM: accumulate expert dW into an existing param.grad by @danielhanchen in #14480 Studio: trim the runtime update card to one summary line and two actions by @oobabooga in #14436 Studio: keep lone asterisks and backticks in finished chat replies by @NilayYadav in #14391 Studio: picking a preset no longer overwrites a model's remembered sampling by @danielhanchen in #14128 FP8 block dequant: ragged / rectangular e4m3fn blocks take the byte decoder on every NVIDIA GPU (4-16x) by @danielhanchen in #14450 Studio: let the owner share a connection with every account by @danielhanchen in #14365 Studio diffusion: allow the multi-GPU layer split on discrete ROCm cards by @danielhanchen in #14479 Studio: make Back in the browser panel work like a browser's by @NilayYadav in #14283 Studio: same diffusion pixels on every B200 server for FLUX.2 Klein and Z-Image by @danielhanchen in #14483 Studio: per-user temp folders, tighter tool sandbox grants, no UNC paths for managed accounts by @danielhanchen in #14461 Studio: MCP widget approval cards show the arguments being sent by @danielhanchen in #14458 Unsloth Studio / Desktop: keep floating panels where they were left across resizes, minimize and restarts by @danielhanchen in #14355 Studio: view, edit and turn off chat memory in Settings > Chat by @danielhanchen in #14357 Studio: login limiter and keyless access behind a reverse proxy by @danielhanchen in #14459 Studio: make CPU RAG embedding fast on Windows ROCm torch and honour Windows CPU affinity by @danielhanchen in #14354 Studio: media decodes open no protocols, Data Recipe names stay inside their folder by @danielhanchen in #14463 Train vision and audio models from unsloth train by @NilayYadav in #14389 Tests: real-accelerator skip guards and utf-8 read in the gpt-oss and gc-skip tests by @danielhanchen in #14504 Block compile: test that checkpoint early-stop does not demote the layer by @indrajeetapache in #14065 Keep has_internet from setting the process-wide socket timeout by @vineethsaivs in #14378 Dequantize FP8 weights in the activation dtype for fp16 LoRA by @danielhanchen in #14495 Cross entropy: 16384 / 8 warp chunked forward on T4 and A100 by @danielhanchen in #14498 GRPO: opt-in rollout logprobs as the old policy (UNSLOTH_GRPO_ROLLOUT_LOGPROBS_AS_OLD) by @danielhanchen in #14489 Studio diffusion: keep a model-offloaded denoiser's step eager instead of re-recording its CUDA graph every render by @danielhanchen in #14497 Studio: route krea2 GGUFs to Images instead of opening them as chat models by @indrajeetapache in #14477 Copy uv constraint files to a space-free path on Windows volumes without 8.3 names by @danielhanchen in #14507 Fix main's red CI after #14437, #14455 and #14473 by @danielhanchen in #14517 CLI: --parallel help describes the shared KV cache by @danielhanchen in #14340 FSDP2 gloo tests: stub every helper the trainer wrapper calls, not a fixed list by @danielhanchen in #14518 Windows installer: restore the split-layout comment in setup.ps1 (Defender ML false positive) by @danielhanchen in #14520 Tests: follow #14473's Unsloth wording in the sandbox, xformers and autoopt parity checks by @danielhanchen in #14522 Add REAP expert pruning for MoE models by @danielhanchen in #13834 Studio: opt-in REAP expert pruning on the Export page by @danielhanchen in #14130 Studio: send the TTS language to Supertonic and add a Speak language picker by @danielhanchen in #14511 unsloth start openclaw: stop writing the legacy exec-approvals.json that OpenClaw 2026.8.1+ refuses to run with by @danielhanchen in #14343 Studio: whisper.cpp updates ride along with llama.cpp instead of opening the update card by @danielhanchen in #13862 QAT: qat_scheme = "nvfp4", and strip NVFP4 / MX QAT linears before saving by @danielhanchen in #14512 Studio diffusion: key learned peaks on the unsloth mirror a fresh install downloads by @danielhanchen in #13847 Dequantize fine-grained fp8 checkpoints on a 16bit load in FastLanguageModel, and save them as 16bit by @danielhanchen in #14516 Studio diffusion: place bf16 and GGUF / int8 packed denoisers from their learned peak by @danielhanchen in #13860 Load FP8 checkpoints dequantized to 16-bit on GPUs below sm89 by @danielhanchen in #14531 Cast float32 activations into the float16 projections of Falcon-H1 / Nemotron-H fp16 loads by @danielhanchen in #14532 NVFP4 fused dX: skip split K on sm89 when the output already fills the GPU by @danielhanchen in #14536 MoE grouped GEMM: share autotuners across identical cards; fix the TMA test on one-GPU hosts and macOS by @danielhanchen in #14490 Jetson / Rosetta follow-ups: an explicit GPU index beats the Jetson CPU record, JetPack 7 off the CPU note by @danielhanchen in #14494 Studio diffusion: compile group-offloaded blocks below their hooks on the deferred profile and for planned-offload video by @danielhanchen in #14528 Wait for the Unsloth-managed folder label in the project folders browser smoke by @danielhanchen in #14542 Studio: Z-Image guided steps run at the guidance-off batch so turning CFG on recompiles nothing by @danielhanchen in #14541 Keep two Backend CI tests valid on uv 0.13 and on runners with no accelerator by @danielhanchen in #14535 Format the grouped GEMM single-flight wrapper so pre-commit.ci passes again by @danielhanchen in #14544 Studio: refuse a single-file diffusion load that would exhaust host RAM instead of crashing by @danielhanchen in #14510 XPU bf16 on the current device, keep explicit bf16 on XPU, MLX full finetuning keeps its default dtype by @danielhanchen in #14493 Load the real loader_utils helpers into the precision conflict test harness by @danielhanchen in #14545 Studio: FLUX.3 upsampler pin follows torch's vendor, single-GPU hosts launch as before, non-blocking Vulkan probe by @danielhanchen in #14491 Studio: discrete-first GGUF placement keeps explicit device picks, matching context ceiling; in-process Rosetta check by @danielhanchen in #14492 MoE grouped GEMM: per-expert dW kernel by @danielhanchen in #14550 Studio Images: paste generation parameters, inpaint mask blur and only-masked area, seed buttons by @danielhanchen in #14508 Count the context ceiling's per-device reserve gate from #14492 in the compute buffer wiring test by @danielhanchen in #14560 MoE autotune: run the kernel tuning eagerly when called from a compiled block by @danielhanchen in #14559 Studio: Z-Image CFG split under static step skip and CFG-parallel by @danielhanchen in #14551 Take the least of three RssAnon peaks in the block dequant host RAM check test by @danielhanchen in #14561 Studio: show the AMD NPU by default again, with a clearer setup notice by @danielhanchen in #14524 Studio Images: per-generation Shift for flow-matching models and SDXL textual inversion embeddings by @danielhanchen in #14563 Pin Lemonade 2026.43.3 by @danielhanchen in #14567 Studio: pin the master-982 stable-diffusion.cpp prebuilt so MOSS-Image GGUFs load by @danielhanchen in #14565 Give the startup transfer budget headroom again after #14524 set it to its own build by @danielhanchen in #14571 Skip the compiled MoE autotune test on the recorded real CUDA, not the spoofable probe by @danielhanchen in #14577 MoE autotune cache: compare device capability as a tuple so saved configs are reused by @danielhanchen in #14578 Measure the block dequant peak in a fresh process by @danielhanchen in #14579 MoE fallback kernel configs: fit the device shared memory for float32 operands by @danielhanchen in #14576 Studio diffusion: keep fp8 single-file weights stored in fp8 where widening would offload or crowd host RAM by @danielhanchen in #14549 Studio: keep CFG-parallel on the default speed tier and run Z-Image's guidance branches on two cards by @danielhanchen in #14564 Studio diffusion: turn on the measured-safe optimizations by default (fused FLUX RoPE, CFG batching, Blackwell native attention, BF16 small DiTs) by @danielhanchen in #14566 Studio video: export GIFs with one clip-wide palette (about 3x faster) by @danielhanchen in #14581 Studio diffusion: latch failed fused fp16 RoPE and AdaLN launches for the load by @danielhanchen in #14583 RoPE cache: build cos/sin outside inference_mode so GRPO keeps the compiled decoder block by @danielhanchen in #14584 Keep Unsloth Studio data after docker rm on CPU-only Docker hosts by @NilayYadav in #14540 Fast decode: rotate Q and K in one launch, and fuse Qwen3 q/k norm with RoPE by @danielhanchen in #14599 Honor requested languages for Studio Qwen3-ASR transcription by @tompipe in #14548 Studio: show relative and blob Hugging Face images in Hub model cards by @shimmyshimmer in #14602 Studio: chat fork fixes and clearer fork affordances by @shimmyshimmer in #14587 Studio: Projects recency by chat activity, less polling behind projects and sources by @shimmyshimmer in #14586 QAT: qat_scheme = "mxfp8", and let MX QAT linears take float32 activations under autocast by @danielhanchen in #14607 unsloth start: fix Studio discovery, proxy, Hermes context, Claude hooks and Codex login by @shimmyshimmer in #14591 Studio: faster, clearer model folder management by @shimmyshimmer in #14588 MoE grouped GEMM: no lock on cached autotuner launches; read the SM count once per card by @danielhanchen in #14605 Studio: token activity shows a year of quartile-shaded days, with accurate counts behind it by @shimmyshimmer in #14600 Studio: shortcut fixes for non-US layouts, Esc and Cmd-K, plus clearer settings by @shimmyshimmer in #14597 FLUX.3 upsampler: name older NVIDIA Vulkan devices as NVIDIA, probe an empty answer once per window by @danielhanchen in #14604 Studio: more accurate connection checks and clearer errors by @shimmyshimmer in #14594 Studio Run settings: restore saved settings on a fresh editor, show save state, group Advanced rows by @shimmyshimmer in #14589 Fast LayerNorm backward: no aliased gradient buffers when only one of weight / bias trains by @danielhanchen in #14609 Fused decode RoPE: drop the per-layer host drain on split models by @danielhanchen in #14611 Studio: resize the sliding pill with transforms, so it keeps pace with the move by @Etherll in #14361 Studio: OpenAI-compatible fine-tuning jobs API (/v1/files, /v1/fine_tuning/jobs) by @danielhanchen in #14496 MoE autotune cache: version disk entries and refit loaded configs by @danielhanchen in #14617 Studio Images: SDXL prompt weighting, CLIP skip, CFG rescale and variation seeds by @danielhanchen in #14562 NF4 bitsandbytes GEMM gate: return before its lazy setup under torch.compile by @danielhanchen in #14608 Studio: say when an attached image is over the 20 MB limit by @NilayYadav in #14538 Studio: accept skills whose description has an unquoted colon by @NilayYadav in #14537 Studio: keep Decision API models out of the chat model picker by @NilayYadav in #14580 Studio: refuse Laya choice questions whose options would be cut short by @NilayYadav in #14558 Give the single-row cached attention harness CPU stand-ins for the fused decode RoPE kernels by @danielhanchen in #14632 Studio: shortcut follow-up to #14597 for Swiss/Turkish digit chords and Esc-to-stop by @shimmyshimmer in #14633 Studio: busiest day keeps the darkest token activity shade (follow-up to #14600) by @shimmyshimmer in #14630 Studio: keep Projects page renames and deletes from waiting out a streaming reply by @shimmyshimmer in #14634 Studio: keep folder identity lookups inside the scan budget (follow-up to #14588) by @shimmyshimmer in #14638 Studio: name an untitled chat's fork after its first reply (follow-up to #14587) by @shimmyshimmer in #14636 unsloth start: hard-link only auth.json into the Codex parent overlay (follow-up to #14591) by @shimmyshimmer in #14635 Studio: default scheme-less public provider URLs to https (follow-up to #14594) by @shimmyshimmer in #14629 NVFP4 fused GEMM: honor autocast dtype and check K alignment for dX by @danielhanchen in #14625 MoE LoRA: widen gate_up_proj to split gate_proj / up_proj experts when down_proj already matches by @danielhanchen in #14628 Studio: engine versions and updates in Settings, Use via API for Images and Video, shorter Notifications, rounder waveforms by @shimmyshimmer in #14590 Studio Settings: stop forcing a GitHub release lookup on every Settings open by @shimmyshimmer in #14649 Studio: stop long MLX chats from re-reading the whole chat every turn by @NilayYadav in #14574 Studio diffusion: test First Block Cache on a two-card FLUX.2 layer split by @danielhanchen in #14620 Studio: load GGUF models faster by skipping the vocabulary read by @NilayYadav in #14573 Fix save_pretrained_torchao for full finetuning QAT on transformers 5 by @danielhanchen in #14639 RMSNorm: enable the narrow-row kernels on ROCm with the DPP reduction order by @danielhanchen in #14637 Studio: stop a Clef decision model from outliving a crashed Unsloth by @NilayYadav in #14556 Unsloth Studio: serve an MLX model through the library that loaded it by @Lyxot in #14612 Studio composer icon check: re-measure a case when the dev page reloads mid-case by @danielhanchen in #14647 Trim comments in the Studio frontend tests by @danielhanchen in #13859 Trim comments in the Studio backend tests by @danielhanchen in #13874 Studio: open localhost addresses typed in the desktop browser panel by @NilayYadav in #14290 Gate the fused decode RoPE tests on a real CUDA device, not the spoofable probe by @danielhanchen in #14661 Windows UI lane: retry the permission suite's auth wipe while the stopped server releases auth.db by @danielhanchen in #14658 Studio: speak through a saved connection on the Audio page by @NilayYadav in #14288 Native FSDP2 float16: fix fused AdamW crashing at the first step by @danielhanchen in #14663 Chat: the composer dock stops searching the whole thread on every com… by @Sneakr in #14585 Studio: keep the image info stamp off by default, recipe stays in the metadata by @danielhanchen in #14669 Studio: cut polled-probe cost and release memory sooner by @shimmyshimmer in #14595 Studio: stop showing a crashed chat model as loaded by @NilayYadav in #14575 feat: add EPUB and optional MOBI dataset loading by @HughScott2002 in #14453 Studio: optional reply times in Settings > Chat by @danielhanchen in #14500 Studio: unload an idle decision model when another model loads by @NilayYadav in #14555 FastSentenceTransformer: non-reentrant gradient checkpointing under DDP for encoder models by @danielhanchen in #13639 Studio: stop the SDPA probe reporting fused attention on MPS by @Lyxot in #14488 Studio: extract ebooks, Office files and email in Data Recipes document seeds by @danielhanchen in #14499 FP8: keep every scoped converter and clean up every dequantizing load before save by @danielhanchen in #14678 Attention routing tests: follow torch's own GQA kernel choice (fixes two failures on torch 2.14) by @danielhanchen in #14672 Studio: lighter system polling and clearer unified-memory monitor by @shimmyshimmer in #14592 Studio settings: fix API, Connections, Remote access and LAN issues by @shimmyshimmer in #14596 Studio: lower the VRAM budget floor to 10% for multi-GPU spreading by @Shooter3k in #13427 MoE grouped GEMM: keep the legacy dW kernel when an explicit tile overflows per-expert shared memory by @danielhanchen in #14723 Studio: warn when a GGUF is tight for Apple unified memory by @shimmyshimmer in #14598 Studio: transcribe audio through a saved connection on the Audio page by @NilayYadav in #14289 Studio: let the owner turn off the Decision API while a model loads by @NilayYadav in #14557 tests: restore transformers' Fp8Dequantize after the fp8 save policy tests by @danielhanchen in #14741 FP8 dequant: multiply per-tensor and rowwise scales in fp32 for fp16 output by @danielhanchen in #14677 Studio: upgrade a Desktop Chat only install to the full install from Settings by @oobabooga in #14593 NVFP4: keep fast decode callers in the activation dtype under autocast by @danielhanchen in #14750 feat(studio): MLX PLE and expert offload in Unsloth Studio by @Lyxot in #14326 Unsloth Studio (Vulkan): fill the discrete GPU before a shared-memory iGPU in a mixed pin by @LeoBorcherding in #12650 CFG-parallel: treat a changed Python number, storage offset or autocast as a cold call by @danielhanchen in #14614 Studio diffusion: keep CFG batching off for a resident denoiser with no device timer (CPU, MPS) by @danielhanchen in #14652 Images: refuse Flux base / max shift values that build NaN sigmas by @danielhanchen in #14613 Studio diffusion: give stored-fp8 denoisers their own planner label instead of dense by @danielhanchen in #14631 Studio sd.cpp: BF16 text encoders by default (Q8_0 / Q4_K_M changed pixels), opt-in lower precision by @danielhanchen in #14697 NVFP4 QAT: require in_features divisible by 32 by @danielhanchen in #14703 Studio diffusion: report compile and CUDA graph engagement in the image and video status by @danielhanchen in #14656 Studio diffusion: replay a model-offloaded denoiser's CUDA graph across renders and pin torchao int8 host weights by @danielhanchen in #14666 Studio diffusion: arm FLUX.1 prompt-length dynamic shapes so int8 does not recompile on a long prompt or new size by @danielhanchen in #14689 FSDP2 QLoRA: keep 4-bit merge operands on one device in push_to_hub_merged by @danielhanchen in #14724 Studio diffusion: arm Z-Image's prompt-length inputs for an int8 / fp8 compile by @danielhanchen in #14715 Block compile: let an enclosing torch.compile own wrapped decoder layers by @danielhanchen in #14704 Studio sd.cpp: one CPU thread per usable physical core on Linux by @danielhanchen in #14694 Studio: clamp the top_k a model passes to search_knowledge_base by @roydonsequeira in #14222 Studio NVFP4: keep ungated and gate-failing video builds out of Auto, check flashinfer against the gate record by @danielhanchen in #14659 Compiled blocks: packed GatedDeltaNet op, flex main kernel in training blocks, PrefixGrouper calls skip the block (0 graph breaks) by @danielhanchen in #14782 4-bit decode: no recompile from the eager setup, one output dtype under autocast, refresh the cached LoRA cast by @danielhanchen in #14754 MoE LoRA targets: keep a requested router sibling when attention modules are also listed by @danielhanchen in #14745 Two-GPU decode RoPE test: check the copy bitwise and the rounding at 2 ulp by @danielhanchen in #14752 Studio Images: Hires fix and ESRGAN model upscalers by @danielhanchen in #14651 Studio diffusion: change baked LoRA subsets and weights without recompiling or recapturing by @danielhanchen in #14692 Studio: build the CFG-parallel replica outside inference mode and catch in-place edits of inference weights by @danielhanchen in #14626 perf(studio): prefill MLX loads in the planned step while experts are on disk by @Lyxot in #14466 tests: keep the fused-NVFP4 env in place after the split-reduction hook override by @danielhanchen in #14789 FP8 rowwise decode: one multi-row GEMV launch for 2 to 16 tokens by @danielhanchen in #14791 FP8 block decode: fused act_quant and narrow-N GEMM for 1 to 16 tokens by @danielhanchen in #14792 RMSNorm kernels: do not specialize on the row count and row stride by @danielhanchen in #14786 MoE autotune cache: pick kernel configs for the operands' GPU by @danielhanchen in #14751 Studio: show a checkpoint's training loss when that step also evaluated by @MohammadHijjawi97 in #13879 Studio Images: run Krea 2 GGUFs on stable-diffusion.cpp on CPU-only hosts by @danielhanchen in #14627 Studio diffusion: Automatic multi-GPU placement fixes and a placement decision in status by @danielhanchen in #14657 Studio: size auto-quantised text encoders at the hosted file in the image picker by @danielhanchen in #14735 GRPO routing replay: prime router formulas before compile and score the PEFT reference without replay by @danielhanchen in #14756 MoE dW accumulation: fall back when autograd Node hooks can rewrite dW by @danielhanchen in #14759 Gemma 4 vLLM fallback: check the engine's GPU, not the current one by @danielhanchen in #14778 Fused preference loss: fall back when lm_head has module backward hooks by @danielhanchen in #14788 Studio Images: Compare grid (X/Y/Z plot) by @danielhanchen in #14624 MXFP4 packed Linear: bound the below-sm_80 weight decode by free memory by @danielhanchen in #14784 Studio diffusion: one host read per guided step in the CFG-batch and CFG-parallel argument checks by @danielhanchen in #14734 Studio Images: detect v-prediction SDXL fine-tunes from their own config, not their name by @danielhanchen in #14746 Studio: resolve torchao's int8 GEMM home before the int8 fused install returns by @danielhanchen in #14771 Studio diffusion: read ROCm from the wheel tag in every fast-path gate by @danielhanchen in #14654 Studio diffusion: seed dense fp16 denoisers on the default and max tiers from measured peaks by @danielhanchen in #14720 Studio video: stream-encode Wan clips under the ROCm bf16 VAE decode by @danielhanchen in #14653 MoE grouped GEMM: skip kernel-info logging work unless debug logging is on by @danielhanchen in #14787 Studio MLX images: tile the VAE decode when the untiled one would not fit by @danielhanchen in #14796 QAT: fake quantize Gemma embeddings on the float32 and Accelerate-hooked paths by @danielhanchen in #14702 FP8 block decode: weight-only GEMV where it beats the block GEMM, fused small-M called directly by @danielhanchen in #14813 Studio MLX images: convert Q4_K / Q5_K / Q6_K and Q5_0 / Q5_1 GGUFs by @danielhanchen in #14794 Studio MLX images: plan text encoder and dense bits from the Metal working set by @danielhanchen in #14793 Studio: cap Pillow below 13 for the MLX image engine by @danielhanchen in #14802 Studio tests: pin the card name in the learned-peaks store fixture by @danielhanchen in #14804 Studio diffusion: fast whole-component swap for low_vram offload by @danielhanchen in #14767 Fused LoRA: fall back to PEFT for variants, disabled or merged adapters, trainable bases and autocast dtype mismatches by @danielhanchen in #14806 Attention: let the SDPA GQA probe trace under torch.compile(dynamic=True) by @danielhanchen in #14779 FSDP2 NF4 saves: stage packed bytes on the CPU, restore QuantState on full DCP load by @danielhanchen in #14739 Studio tests: compare MLX snapshot paths as paths, not posix strings by @danielhanchen in #14809 Studio tests: skip the two-card fast pin test on ROCm instead of erroring by @danielhanchen in #14618 Apply the repo formatter to 26 files merged without it by @danielhanchen in #14817 Studio MCP: drop a queued video's request and caller once it settles by @danielhanchen in #14825 Studio images: keep the resident GGUF's base repo on Reapply and native reload by @danielhanchen in #14828 Studio video: keep the stream encode hook out of a compiled VAE decode graph by @danielhanchen in #14822 Studio: restore a kept media model's price after a failed replacement by @danielhanchen in #14826 Format test_video_stream_encode.py, merged without the repo formatter in #14822 by @danielhanchen in #14852 Studio diffusion: onload torchao int8 / fp8 weights through the pinned chunks by @danielhanchen in #14832 Studio: pick the vision projector precision per device, and never F16 for MiMo-VL by @danielhanchen in #14679 Studio MLX images: list Q5_0 / Q5_1 in the unsupported GGUF refusal by @danielhanchen in #14839 Studio diffusion: hold cudnn.benchmark off for FLUX.2 klein so a seed decodes the same image in every process by @danielhanchen in #14770 Studio: budget HunyuanImage-2.1's automatic int8 text encoder at its hosted file by @danielhanchen in #14833 Studio: hold the image render slot while a cold FLUX.3 upsampler start comes up by @danielhanchen in #14830 Studio: API keys can't make a Data Recipe read credentials from the server environment by @danielhanchen in #14699 Studio: only trust llama-server's /health when our own child holds the port by @danielhanchen in #14706 Fix Gemma 4 E2B / E4B DPO CheckpointError after the ref adapter switch by @danielhanchen in #14705 Studio: fetch VLM dataset row image URLs through the SSRF-guarded fetch by @danielhanchen in #14701 GPT-OSS template: keep earlier answers on the final channel when the turn has thinking by @danielhanchen in #14700 Studio MLX images: release MLX's buffer cache after each generation by @danielhanchen in #14795 Studio diffusion: keep the CFG-parallel weight probe inside large inference weights by @danielhanchen in #14844 Studio Images: run the Hires fix latent encode where an offloaded SDXL VAE computes by @danielhanchen in #14851 Studio: read the prequant source etag from the repo blob under the shared Hub blob store by @danielhanchen in #14819 Studio updater: install a matched unsloth + unsloth-zoo pair when the index lags by @danielhanchen in #14834 Desktop: keep the environment when WEBKIT_DISABLE_COMPOSITING_MODE=0 leaves compositing on by @danielhanchen in #14837 Studio browser: keep a temporary chat's hidden pages out of history by @danielhanchen in #14855 Studio diffusion: latch the CFG-batch verdict per load and across processes so a seed keeps one path by @danielhanchen in #14849 Format two files merged without the repo formatter in #14834 and #14819 by @danielhanchen in #14878 Studio: account management needs a UI session by @danielhanchen in #14872 Studio diffusion: free the failed group offload setup before the OOM retry by @danielhanchen in #14821 Studio: Install Training button in the Train hint and the Export banner by @danielhanchen in #14861 Tests: name text encodings and gate CUDA skips on has_real_cuda by @danielhanchen in #14871 Studio diffusion: compile MiniMax-H3 with torchao float8 weights by @danielhanchen in #14621 Studio H3: let Auto residency see the card again on ROCm and NVIDIA without nvidia-smi by @danielhanchen in #14820 Studio diffusion: pool SDXL textual inversion prompts at EOS by @danielhanchen in #14846 Studio: flag a modules.json folder of only dots and slashes that leaves the model by @danielhanchen in #14527 Studio: a mermaid style statement after a semicolon or mid-line cannot fetch either by @danielhanchen in #14526 Studio: refuse a chained dataset media path whose first hop is not an archive by @danielhanchen in #14529 Studio: chat audio's librosa fallback refuses playlists too by @danielhanchen in #14530 Studio: a managed account's own Hub token covers a resident loaded from the Hub cache by @danielhanchen in #14533 Studio: read every forwarded header field behind a trusted proxy by @danielhanchen in #14534 Studio: approved host calls keep chat-installed packages, helper modules and CLIs by @danielhanchen in #14543 install.sh: repair a stale ROCm-on-WSL drop-in without an interactive password prompt by @danielhanchen in #14641 Studio: let the browser agent follow search-result redirect links shown on the page by @danielhanchen in #14642 Desktop: write the Windows login entry for the current user only by @danielhanchen in #14643 Studio: Data Recipe column names with a slash run again; the validator file name is made safe instead by @danielhanchen in #14644 Studio: registered model folders and HF cache on /run/media drives stay silent for the owner by @danielhanchen in #14646 Desktop: reject the compromised eMudhra/emSign intermediates (rustls-webpki 0.103.16) by @danielhanchen in #14879 Studio and Core: certifi 2026.10.19 and pip 26.2.2 floors (drop the compromised emSign root) by @danielhanchen in #14887 Studio diffusion: store request peaks for Wan pipelines with an empty transformer_2 slot so measured residency runs by @danielhanchen in #14864 FP8 per-tensor / rowwise dequant: one Triton pass over the e4m3fn bytes by @danielhanchen in #14886 Format two files merged without the repo formatter in #14821 and #14886 by @danielhanchen in #14885 Studio NVFP4: tag the compile cache key with the one-pass activation quantiser so a warm cache cannot bypass its switch by @danielhanchen in #14853 Studio sd.cpp: load a cached earlier text encoder quant when the BF16 default cannot be fetched by @danielhanchen in #14862 Studio: use a cached earlier FLUX.3 upsampler quant when the default is missing by @danielhanchen in #14842 Studio: install the dataset image URL guard when datasets loads, not at startup by @danielhanchen in #14888 Studio: MOSS-Image (in-tree port, quantised builds, GGUF through stable-diffusion.cpp) by @danielhanchen in #14552 Studio diffusion: stop flushing the CUDA allocator on every prompt encode under block offload by @danielhanchen in #14777 Studio sd.cpp: stage server-mode image LoRAs as hard links so sd-server lists them by @danielhanchen in #14860 Studio diffusion: stop recording CUDA graphs once the VMM arena has failed, and detach it by @danielhanchen in #14867 Keep Gemma 3n AltUp coefficients out of on-the-fly 4-bit quantization by @danielhanchen in #14729 Load Qwen3-ASR checkpoints saved in the qwen-asr package layout by @danielhanchen in #14716 Fine-tune Whisper, Canary, Moonshine, CTC, audio classification, Dia and Higgs Audio through FastModel by default by @danielhanchen in #14744 Load VibeVoice-ASR, MusicGen, Audio Flamingo 3 and Music Flamingo through FastModel by @danielhanchen in #14781 XPU bf16: judge the XPU in use, not every visible one by @danielhanchen in #14769 gpt-oss bnb 4-bit: probe the checkpoint transformers loads and keep fused expert quant state on save by @danielhanchen in #14740 Studio Images: Hires fix latent upscale interpolates the base latents directly, as ComfyUI's LatentUpscale by @danielhanchen in #14863 Studio Images: run ESRGAN upscalers through the spandrel descriptor by @danielhanchen in #14876 Tests: rewrite the remote_images import in the no-torch format_conversion harness by @danielhanchen in #14903 Studio: rework the keyboard shortcuts page by @shimmyshimmer in #14896 Security audit: baseline the pandas 3.1.0 find_stack_level test by @danielhanchen in #14870 Format two files merged without the repo formatter in #14769 by @danielhanchen in #14908 Studio ROCm: keep stored flash-attention verdicts from recompiling per bucket, and gate the CFG replica too by @danielhanchen in #14869 Fetch get_statistics repos concurrently so from_pretrained does not wait on four HF round trips by @danielhanchen in #14476 FP8 per-tensor / rowwise dequant under torch.compile below sm89: one opaque byte-decode op by @danielhanchen in #14913 Studio: hover label on the composer + button by @shimmyshimmer in #14927 Studio video: fit per-request residency on the streaming tier and for HunyuanVideo-1.5 by @danielhanchen in #14865 Studio diffusion: seed FLUX.2 klein residency from its measured peak on the first load by @danielhanchen in #14866 Studio Images: Compare grid counts Hires CFG when sending the negative prompt and records Hires settings by @danielhanchen in #14875 FSDP2 QLoRA merged saves: stage the packed 4-bit base on the CPU and merge one layer at a time by @danielhanchen in #14823 Studio: steadier run settings controls, grey sliders, 18px thumbs by @shimmyshimmer in #14926 Format four files merged without the repo formatter in #14913 and #14823 by @danielhanchen in #14930 Installers: pin uv 0.13.5 and replace uv older than 0.13.4 (emSign root) by @danielhanchen in #14894 Studio: tell websites the macOS browser panel is Safari by @NilayYadav in #14797 Studio Images: keep v-prediction detection on cache-only SDXL loads by remembering the online markers per snapshot by @danielhanchen in #14873 Studio: horizontally scroll long API usage examples by @NilayYadav in #14798 Studio: list imatrix-weighted quants named --imatrix.gguf by @danielhanchen in #14931 Studio: remove the New badge from the Browser settings tab by @shimmyshimmer in #14939 Large head attention: keep CUDA graph capture off the mask scan and bound int32 row offsets by @danielhanchen in #14790 GRPO: sync LoRA weights to TRL's own vLLM engine on TRL < 0.28, refuse quantized bases by @danielhanchen in #14831 Native FSDP2: keep accelerate FSDP2 for custom Accelerator loops, support unsharded generation on torch 2.6 / 2.7 by @danielhanchen in #14882 Keep the checkpoint_skip_layers="auto" planner out of torch.compile regions by @danielhanchen in #14928 Studio: gate saved Connection keys on a stamped key purpose, not the key name by @danielhanchen in #14824 FP8 block dequant on gfx11 ROCm: ragged / rectangular e4m3fn blocks take the byte decoder by @danielhanchen in #14936 Studio: model picker hides empty uploads and fades models this device cannot run by @shimmyshimmer in #14935 Studio: save sampling defaults per model and use them for API requests by @danielhanchen in #14515 Keep HTML previews visible when requesting edits by @NilayYadav in #14810 FP8 gfx11 byte decoder: tighten comments from #14936 by @danielhanchen in #14947 Studio chat: go back to a chat's model when it is still loaded by @debojyoti-dev in #14568 Studio: one fixed radius scale, applied to the shared primitives by @shimmyshimmer in #14919 Studio: preserve conversation archive order for turns with attachments by @oobabooga in #14768 Fix delayed playback of generated audio by @Imagineer99 in #14671 Studio: raise the chat template limit from 64 KiB to 512 KiB by @danielhanchen in #14934 DiffusionGemma: build the prefix cache per DataParallel replica by @danielhanchen in #14942 Attention: keep packed fp32 on one xformers call, read the fp32 downcast from Q's device by @danielhanchen in #14805 Studio: pin managed dataset media URLs to the Hub and scope extracted archive members by @danielhanchen in #14840 Tests: follow the radius scale in the titlebar and remote access contracts by @danielhanchen in #14957 Check the weight index of dot-named Sentence Transformers module folders by @danielhanchen in #14952 Format dataset_media.py, merged without the repo formatter in #14840 by @danielhanchen in #14963 CodeQL: run weekly and on dispatch only by @danielhanchen in #14953 Studio: reuse Model Hub filter metadata between searches by @alkinun in #14265 fix(export): pass trust_remote_code for checkpoint exports by @ryder777777 in #14722 Studio (Windows): pair a kept torch 2.12 with triton-windows 3.7 by @danielhanchen in #14684 Studio: keep causal-conv1d for Qwen3.5 / Qwen3-Next remote implementations by @danielhanchen in #14685 Studio setup: keep the extension-free mamba-ssm in the stale kernel check by @danielhanchen in #14687 Installer (Windows): fall back to the python-build-standalone mirror for uv-managed Python by @danielhanchen in #14686 Studio (AMD Linux): keep the bundled ROCm runtime when the system ROCm prepend hides the GPU by @danielhanchen in #14933 Fix centering of audio and music results by @Imagineer99 in #14718 Studio: simpler API usage examples, and no Pi / Mistral Vibe outbound links by @shimmyshimmer in #14914 CI: skip Windows, macOS and Tauri jobs on draft pull requests by @danielhanchen in #14971 Studio: tidy the Logs viewer and let it resize by @shimmyshimmer in #14965 Studio diffusion: report the split's primary card as the execution device for diffusers ModularPipeline (FLUX.3 on two cards) by @danielhanchen in #14917 Studio: neutral grey tints instead of green ones by @shimmyshimmer in #14938 Studio: give Intel XPU Triton a working C++ toolchain on Windows by @debojyoti-dev in #14814 Studio: no divider above the Live monitor's Detailed view row by @shimmyshimmer in #14972 Studio: remove unused image assets by @shimmyshimmer in #14967 Pin FastFlowLM v1.0.10 by @danielhanchen in #14978 Studio: keep renamed audio.cpp options at generation time by @danielhanchen in #14901 Studio: drop file toolbar controls that repeat a visible button by @shimmyshimmer in #14976 Studio: trim file toolbar comments by @shimmyshimmer in #14980 Load and QLoRA-train Moonshine Streaming checkpoints by @danielhanchen in #14945 Studio: attach duplicate-named text files like LICENSE(1), and say why a file can't be attached by @shimmyshimmer in #14977 Studio: stop the macOS desktop app typing stray control characters for unhandled keys by @debojyoti-dev in #14815 Unsloth Studio: browser use fixes and clearer agent progress by @shimmyshimmer in #14681 Studio diffusion: CFG-parallel router keeps the wrapped forward's signature (FLUX.3 under CFG-parallel) by @danielhanchen in #14918 Studio: show when a chat is waiting for tool approval by @oobabooga in #14877 Tests: stub _unmerge_flags_on_error in the merged hub destination harness by @danielhanchen in #14961 FP8 e4m3fn scale dequant: pass enable_fp_fusion like every other launch by @danielhanchen in #14969 Studio: dataset image guard keeps owner flows working (opt-in, WSL paths, unencoded spaces) by @danielhanchen in #14966 Studio diffusion: record each card's CUDA graphs on its own capture stream, so compiled renders on a second GPU are correct by @danielhanchen in #14915 Studio: let HTTP/SSE MCP tool calls run for the whole Max Tool Call Duration by @SauravBellam in #14308 Studio: keep turns an external provider's compaction drops searchable by @debojyoti-dev in #14731 Studio: refuse to start Claude Code or OpenCode on a context too small for their own prompt by @NilayYadav in #14799 Studio sd.cpp: keep sd.cpp's all-CPU thread count on small SMT hosts nothing restricts by @danielhanchen in #14847 Studio: fix host memory checks and improve logging for GGUF image loads by @VTplus in #14514 Windows: read safetensors shards with pread on transformers 5.0 - 5.16 by @danielhanchen in #14856 Studio: split Z-Image guided steps only while the denoiser is compiled or replaying graphs by @danielhanchen in #14868 New Contributors @ATAMAH made their first contribution in #14098 @ShryukGrandhi made their first contribution in #14304 @keyiyu0331 made their first contribution in #13668 @cedric66 made their first contribution in #11813 @yejonz made their first contribution in #13987 @SIDDARTHAREDDY8 made their first contribution in #13748 @tompipe made their first contribution in #14548 @Shooter3k made their first contribution in #13427 @roydonsequeira made their first contribution in #14222 @ryder777777 made their first contribution in #14722 @SauravBellam made their first contribution in #14308 @VTplus made their first contribution in #14514 Full Changelog: v0.1.909-beta...v0.1.910-beta

Performance and Compatibility

Performance is one of Unsloth's main purposes. Its training optimizations are designed to reduce memory requirements and speed up fine-tuning compared with conventional training workflows.

Actual performance depends heavily on the model, quantization, GPU, available VRAM, dataset, batch size, and training configuration. The project supports multi-GPU setups and provides specialized guidance for newer NVIDIA hardware, AMD GPUs, Intel GPUs, and Apple silicon.

Unsloth Studio can run on Windows, Linux, WSL, and macOS. The project also provides native desktop packages for Windows, macOS, and Linux, while the Studio interface can be installed separately and accessed through a local web interface.

macOS users can run models through MLX and GGUF, while supported Apple Silicon systems can also perform training workflows. NVIDIA GPUs provide the broadest support for training workloads, while AMD and Intel support varies by backend and feature.

Because Unsloth works with local models, storage requirements can become significant. Model weights, datasets, checkpoints, caches, and exported models can consume substantially more storage than the application itself.

System Requirements

Unsloth does not have one fixed hardware requirement because different workloads have very different resource requirements.

Supported platforms include:

• Windows

• Linux

• WSL

• macOS

• NVIDIA GPUs

• AMD GPUs

• Intel GPUs

• Apple Silicon

• CPU-only operation for supported workloads

For the code-based installation, the current documentation uses Python 3.13 with uv for Linux, WSL, and Windows installations.

GPU training requires a compatible backend and sufficient memory for the selected model and training configuration. Larger models generally require more VRAM or system memory, while quantized models can substantially reduce the memory requirement.

Docker is also supported, with official images available for different environments and GPU configurations.

Pros and Cons

Pros

• Free and open source

• Local model inference

• Fine-tuning support

• Memory-efficient training optimizations

• Supports multiple model types

• GGUF support

• MLX support

• RAG capabilities

• MCP support

• Tool calling

• Code execution

• Multi-GPU support

• Windows, Linux, WSL, and macOS support

• NVIDIA, AMD, Intel, and CPU support

• Apple Silicon support

• Docker support

• Desktop application

• Web interface

• Python-based workflows

• Supports AI coding agents

Cons

• Hardware requirements vary significantly by model

• Large models require substantial VRAM or system memory

• Training workflows can be technically complex

• Different hardware backends do not provide identical capabilities

• Model files and training datasets can consume considerable storage

• Some advanced workflows require command-line or Python knowledge

How to Install

The simplest option is Unsloth Desktop. The project provides native packages for Windows, macOS, and Linux. Linux users can choose DEB or AppImage packages, while macOS and Windows have dedicated installers.

For Unsloth Studio, macOS, Linux, and WSL can use:

curl -fsSL https://unsloth.ai/install.sh | sh

Windows can use:

irm https://unsloth.ai/install.ps1 | iex

After installation, start the Studio interface with:

unsloth studio

The interface runs locally and can be accessed through a web browser.

For developers who prefer the Python package, Unsloth Core can be installed inside a virtual environment. The project currently recommends using uv to create the environment and install Unsloth with automatic PyTorch backend selection.

Docker is another option for users who prefer an isolated environment or need a reproducible setup.

Final Verdict

Unsloth is more than a fine-tuning library. Its current ecosystem combines local model inference, training, model conversion, RAG, tool calling, AI agents, and a graphical Studio interface into a single open-source platform.

Its biggest strength is flexibility. Users can start with the desktop application and local models, then move to Python-based fine-tuning or Docker when they need more control. Support for multiple hardware platforms also makes it useful across a wider range of systems, although the available capabilities vary between backends.

The main limitation is complexity. Unsloth is aimed at users working with local AI models rather than people looking for a simple chatbot. Model size, VRAM, training configuration, and hardware compatibility all have a significant impact on the experience.

Free

Download

Software Informations:
Developer:
Unsloth
Operating System:
Windows / macOS / Linux
Date Added:
2026-10-24T03:07:25.733Z
Categories:

Related to Unsloth v0.1.910-beta:

Post a Comment/Report Broken Link: