Unsloth v0.1.917-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 v0.1.917-beta - Software Mirrors
Unsloth v0.1.917-beta for Windows Unsloth-Desktop-Windows.exe | 20.78 MB Unsloth-Desktop-Windows-ARM64.exe | 20.16 MB
Unsloth v0.1.917-beta for macOS Unsloth-Desktop-MacOS.dmg | 21.28 MB
Unsloth v0.1.917-beta for Linux Unsloth-Desktop-Ubuntu.deb | 22.64 MB Unsloth-Desktop-Ubuntu-ARM64.deb | 22.39 MB Unsloth-Desktop-Linux.AppImage | 169.68 MB
Others Download related to Unsloth v0.1.917-beta Unsloth-Desktop-ARM64.app.tar.gz | 21.04 MB
Unsloth v0.1.917-beta Source Code Unsloth v0.1.917-beta Source code (zip) Unsloth v0.1.917-beta Source code (tar.gz)
Unsloth v0.1.917-beta Release Notes: Qwen3.8-27B is here with Dynamic v3.1 GGUFs, ready to run locally with Unsloth. This release also brings major performance improvements, smoother Glass rendering, revamped code blocks and Thinking UI, HTML/PDF chat exports, and new image generation and coding agent features. Highlights New Qwen3.8-27B Dynamic v3.1 GGUFs. Faster chats: Major performance improvements to streaming, scrolling, menus and dialogs, especially in long conversations. Smoother Glass: Reduced lag while typing, scrolling and generating. Revamped code blocks and Thinking: Independently scrolling, expandable panels with improved syntax highlighting. HTML and PDF exports: Export chats directly from the chat menu. Image queue: Queue upcoming image generations while another is rendering. Claude Code and Codex on MLX: Run coding agents with MLX models through unsloth start. Chat & UI Streaming performance no longer degrades as conversations grow. Smoother scrolling with background images, with or without Glass. Faster menus, dialogs and browser pane loading. Read aloud starts after the first sentence instead of waiting for the entire response. Chat preferences, time format and spell check now sync with your account. Images & Video FLUX.2 and FLUX.3 use 36–43% less VRAM during image decoding at the same speed. Queue image generations from the Images page or API. INT8 and FP8 image models now load on older NVIDIA GPUs, including T4. Fixed black image outputs with FLUX.3. LTX-2.5 FP8 now runs on 24 GB GPUs. Cosmos3-Nano can continue video clips from their first or last frames. Coding Agents & API Run Claude Code and Codex with MLX models on Mac using unsloth start. New Unsloth MCP preset lets coding agents access Unsloth tools. Claude API models without an explicit token limit no longer stop at 1,024 tokens. MLX models now return a proper context-length error instead of an empty response. Training model.generate() is now 2–8x faster on T4 GPUs by default across FastModel models. Video columns are now correctly supported when training vision models. Ulysses context parallelism now supports Kimi-Linear, DeepSeek, MiniMax and Gemma 4 text models. What's Changed Bump install.sh / install.ps1 pins to unsloth>=2026.11.3 by @danielhanchen in #16920 Studio GGUF: keep the stock Q8_0 forward on torch below 2.7 by @danielhanchen in #16911 Studio: merge LoRA and LoKR adapters into ComfyUI-quantized weights by @oobabooga in #16874 Studio: restore WebKitGTK offscreen code fences in batches and on any pointer release by @danielhanchen in #16919 Studio: warm the compile workers and Triton / inductor keys during a load that compiles inline by @danielhanchen in #16918 Studio: pre-seed the float specialisations of an image load's first compiled render by @danielhanchen in #16917 Studio diffusion: reuse the early quant probe's result instead of spawning a second child by @danielhanchen in #16910 Studio: keep Glass fast on WebKit by giving glass surface contents their own layers by @danielhanchen in #16921 Studio: release a streamed encoder's safetensors mapping once its weights are pinned by @danielhanchen in #16914 Studio: load long chat history faster on Linux by mounting parked history in larger steps by @danielhanchen in #16924 Studio: decode escaped spreadsheet text during RAG indexing by @DimitriChrysafis in #16712 Studio: refuse MLX prompts longer than the served context by @Lyxot in #16821 Studio: preserve Unicode across continued XLS strings by @DimitriChrysafis in #16713 Studio: classify MLX repos by name or model-card tag everywhere by @Lyxot in #16883 Studio: blur the glass Settings dialog on its own layer so scrolling it stays cheap by @danielhanchen in #16942 Studio CI: long-chat gate measures every interactive area under every appearance mode by @danielhanchen in #16931 Studio: keep Glass cheap on the Linux app by dropping its live blur except the sidebar's by @danielhanchen in #16957 Studio: stop menus, sidebar and browser pane from restyling the whole chat by @danielhanchen in #16938 Studio: stop WebKit restyling the whole page when dialogs, menus and the browser pane open or close by @danielhanchen in #16961 Studio CI: tighten long-chat gate allowances after the :has() rewrite by @danielhanchen in #16978 Studio: preserve S3 dataset files with colliding names by @DimitriChrysafis in #16830 Studio CI: long-chat gate measures the dialog lock again before failing by @danielhanchen in #16985 Studio: stop dialog open and close from rescanning a long chat for every shortcut by @danielhanchen in #16983 Studio: skip layout of distant messages so the browser pane opens fast on long chats by @danielhanchen in #16986 Studio diffusion: return the heap a finished image load freed to the OS by @danielhanchen in #16934 Scan baseline: re-review the seven unsloth-zoo sites 2026.11.3 moved by @danielhanchen in #16922 CI: run the Playwright WebKit jobs on ubuntu-24.04 while ubuntu-latest is 26.04 by @danielhanchen in #16987 Studio: ship built-in float specialisations so a fresh install's first compiled render traces once by @danielhanchen in #16939 Studio diffusion: faster returning first render and first clip (reuse tuned reductions, warm video compile keys) by @danielhanchen in #16981 Inline images: give a slow sandbox image load a bounded grace and say why it timed out by @danielhanchen in #16923 Diffusion route tests: pin a compiling host for the tests that expect fp8 to be admitted by @danielhanchen in #16993 Studio sd.cpp: image render failures carry sd-server's own output by @danielhanchen in #16815 Studio: show long thinking in a capped box that scrolls on its own instead of Expand thinking by @shimmyshimmer in #16876 Studio: keep compacted turns searchable when the archive embedder fails by @debojyoti-dev in #16718 Studio: realtime option for speed and context stats, shorter Chat defaults by @shimmyshimmer in #16888 Studio RAG: embed the documents of one ingest together instead of one call per document by @debojyoti-dev in #16727 Studio images: render on the NVIDIA card when torch drives no GPU by @debojyoti-dev in #16853 Studio CI: long-chat gate Glass ceilings back to 10x (areas) and 18x (Settings scroll) by @danielhanchen in #17003 Studio: quantise the hosted int8 leftovers on the GPU on offloaded loads by @danielhanchen in #17002 Studio: start a first video load's cache writes after its first render by @danielhanchen in #17001 Studio: let unsloth start attach to a loaded Ollama model by @NilayYadav in #16944 Studio: show DeepSeek Harness as installed on Windows by @NilayYadav in #16954 Studio: stop OpenAI chats failing on earlier replies sent as parts by @NilayYadav in #16945 Studio: keep messages from another window when deleting in a stale one by @NilayYadav in #16926 Studio: size the popular audio.cpp TTS and music models from measured peaks on each runtime backend by @oobabooga in #16984 Studio: block more metadata host spellings for approved Python by @NilayYadav in #16937 Studio: offer ChatGPT Reconnect for reused or revoked refresh tokens by @NilayYadav in #16946 Studio: Agents tab says unsloth start claude needs a GGUF model by @NilayYadav in #16952 Studio: stop Export all chats from saving an empty or stale backup by @NilayYadav in #16930 Studio CI: guard the long-chat gate Glass ceilings against runner noise by @danielhanchen in #17012 Studio: auto-start the server from the same install as unsloth start by @NilayYadav in #16948 Studio: show an error when ChatGPT sign-in is denied by @NilayYadav in #16950 Studio: cap chat code blocks at a scrollable height and keep them coloured while scrolling by @shimmyshimmer in #16865 Studio: stop cutting Claude answers at 1,024 tokens without a limit by @NilayYadav in #16953 Studio: keep Codex AGENTS.md, skills and rules in unsloth start codex by @NilayYadav in #16947 CI: key the Playwright browser caches on the OS release, and move the WebKit jobs back to ubuntu-latest by @danielhanchen in #17009 Studio: reload an idle-unloaded GGUF with its original launch settings by @NilayYadav in #16927 Studio: keep the base URL query string on connection requests by @NilayYadav in #16951 Studio: key the modal pointer-lock rules off attributes that never sit above the thread by @danielhanchen in #17006 Studio: keep same-second fine-tuning jobs in submission order across restarts by @oobabooga in #16835 Studio: accept zstd-compressed requests from the TypeSafe SDK by @NilayYadav in #16929 Studio: keep Hermes MCP servers in unsloth start hermes by @NilayYadav in #16949 Studio: report the LTX-2.5 encoder worker host memory on macOS and Windows too by @danielhanchen in #17014 Studio: pick which LoRA to add from a searchable list instead of cycling through them by @oobabooga in #16964 Run Claude Code and Codex on MLX models by @Lyxot in #16811 Studio: hide the Model hub list under an open model page by @danielhanchen in #17021 Tests: let the install damage and reinstall hint tests run on Windows by @danielhanchen in #17022 Studio: keep search icons sharp under the Glass theme on non-chat pages by @debojyoti-dev in #16989 Studio: keep the first in-loop cut for later tool rounds in chats without a saved thread by @NilayYadav in #16936 Unsloth Studio: keep tabs signed in on plain-HTTP origins when two refresh at once by @LeoBorcherding in #16755 Studio: keep chat, time format and spell check settings per account by @NilayYadav in #16958 Unsloth Studio: shorter new-architecture dialog, one Update button, custom code under Other by @LeoBorcherding in #16966 Studio: refuse sandbox connects to cloud metadata addresses at runtime by @danielhanchen in #17024 Unsloth Studio: LFM3 spoken replies and read aloud on the transformers backend by @LeoBorcherding in #16972 Studio: switch light and dark without starting a transition on every element under Reduce motion by @LeoBorcherding in #16977 Unsloth Studio: read aloud starts speaking after the first sentence instead of after the whole reply by @LeoBorcherding in #16955 Unsloth Studio: AMD: skip LFM3's Triton conv and flex attention on ROCm cards Triton cannot build for by @LeoBorcherding in #16974 fix(studio): heal tool calls in the model's own format on the in-process route by @Lyxot in #16824 Unsloth Studio (AMD): retry an unpinned llama-server arch crash on the cards the build has kernels for by @LeoBorcherding in #16513 Block Swap blocker: offload_layers never engages on Qwen3.8-Flash-Next, the CPU-kept n-gram table makes the map look like two devices by @LeoBorcherding in #15633 Studio: memory-map an oversized load on an integrated GPU and read its lookup tables from disk by @oobabooga in #16857 Studio: answer shortcut availability checks from registered surfaces, not the whole document by @danielhanchen in #17005 Unsloth Studio: Cosmos3 guardrail refusal names the gated nvidia/Cosmos-1.0-Guardrail repo by @LeoBorcherding in #16742 Studio: stream the LTX-2.5 text encode on the card under a VRAM budget, no per-prompt encoder copy back by @danielhanchen in #17000 Unsloth Studio: show why a video generation failed instead of only "Video generation failed." by @LeoBorcherding in #16744 Tests: update two contract tests that #16953 and #16927 left red on main by @danielhanchen in #17040 Unsloth Studio: train and export LFM3-Audio-2B (LoRA, merged 16-bit, GGUF) by @LeoBorcherding in #16973 Studio diffusion: race the unlisted-arch int8 GEMM tiles held under load (RTX 3090 class: Qwen-Image-2.1 -10%, MiniMax-H3 -9%) by @danielhanchen in #17026 Studio: name the Cosmos guardrail module outright so the risky-loader lint passes on main by @danielhanchen in #17041 Tests: let the RAG re-embed and embedder alias tests run on macOS and Windows by @danielhanchen in #17037 Unsloth Studio / Desktop: the Images page Compare grid, for API callers (server-rendered X/Y grid, no UI change) by @LeoBorcherding in #15169 Studio: Live monitor with CPU, memory and per-GPU rows, Unsloth's share by model, and a refresh rate by @LeoBorcherding in #15630 Studio int8 GEMM: keep sm8x GPUs on cuBLAS under Triton 3.8+, where every shipped tile spills by @danielhanchen in #17036 Studio: auto runs compiled int8 W8A8 for Qwen-Image families on RDNA by @danielhanchen in #17013 Studio int8 GEMM: read and write the tile race cache as UTF-8 by @danielhanchen in #17044 Refuse fp8 training layers under tensor parallel instead of silently training in 16 bit by @danielhanchen in #16881 Context parallel: Unsloth's fast Mistral, Granite and Cohere attention, and DPO / GRPO on fast Qwen3 by @danielhanchen in #16893 Studio tests: drain deferred work before the auth refresh fixture restores window by @danielhanchen in #17055 TP x FSDP2 save: one writer when only tp-mesh DTensors remain by @lonexreb in #15257 Tensor parallel sharded loading: allow init_lora_weights = "orthogonal" by @lonexreb in #15218 ORPO / CPO / DPO / KTO chunked head: fall back under torch's global module hooks by @lonexreb in #15259 Studio diffusion: decode image VAEs into one expandable allocator segment (SDXL holds 1.15 GiB less VRAM) by @danielhanchen in #17045 FSDP2 sharded loading: allow orthogonal LoRA init by @danielhanchen in #17038 PPO / RLOO fused log-probs: fall back under torch global module hooks by @danielhanchen in #17043 Context parallel: Ulysses for Kimi-Linear (KDA + MLA), and fix Kimi-Linear MoE gate backward by @MegaGimen in #15880 GRPO: score a hooked output head through its own forward by @danielhanchen in #17050 Tests: skip the compiled half of the fp8 codes layer test on Windows without MSVC by @danielhanchen in #17073 Studio diffusion: Automatic CFG-parallel no longer refuses an idle card over its own CUDA context by @danielhanchen in #17057 Context parallel: register the Qwen3.5 hybrid cp x native FSDP2 cell, add coverage tests by @danielhanchen in #17039 Studio: offloaded native int8 denoisers release their safetensors file once pinned (LTX-2.5 holds ~23 GiB less host mapping) by @danielhanchen in #17065 Studio tests: keep two Linux-host assumptions off the Windows and macOS legs by @danielhanchen in #17107 Studio tests: skip pre-quant tests without torchao and pin the Mac memory reading by @danielhanchen in #17117 Studio MiniMax-H3: run int8 / fp8 denoisers where torchao's runtime does not (Ampere fp8, ROCm, MPS) by @danielhanchen in #17064 Studio: widen stored fp8 / int8 weights without full float32 temporaries by @danielhanchen in #17068 Studio: keep nvfp4 / mxfp8 layers packed when no native runtime runs them by @danielhanchen in #17067 Studio sd.cpp: pass the Windows core count to MiniMax-H3 when sd.cpp would halve it by @danielhanchen in #17089 Studio: AUTO W8A8 on ROCm RDNA keeps bf16 when the int8 plan still offloads by @danielhanchen in #17104 Context parallel: Ulysses for MiniMax (M1 Lightning hybrid, M2) by @MegaGimen in #15881 Studio sd.cpp: pick the Q8_0 BF16 cuBLAS route from the diffusion GGUF, not its companions (1.57x on A100) by @danielhanchen in #17098 Backend CI: retry the ruff-unsloth wheel install on a release-asset 504 by @danielhanchen in #17130 Studio tests: keep the multi-GB fake checkpoints sparse on Windows by @danielhanchen in #17119 Studio diffusion tests: plan as Linux does whatever the host can compile by @danielhanchen in #17135 Studio tests: import the new test helpers top-level when a sibling imports the module that way by @danielhanchen in #17139 Studio: run single-file fp8 on torchao under group offload, and keep Qwen-Image-2.1's timestep table in fp32 by @danielhanchen in #17127 Studio: read GPU metadata from the CUDA driver API, not torch.cuda.get_device_properties by @danielhanchen in #17120 Studio video: spool encoded stream chunks instead of holding a second raw clip by @danielhanchen in #17066 Studio MiniMax-H3: reuse the conditioner output when only the seed of a reference / keyframe render changes by @danielhanchen in #17094 Studio desktop: steady floating browser window on drag, resize and float by @shimmyshimmer in #17027 Studio: keep the lora_adapter_metadata strength of a rewritten LoRA file by @danielhanchen in #17088 Unsloth Studio: train video-capable VLMs on a dataset's video instead of dropping it by @LeoBorcherding in #16965 GRPO PrefixGrouper: engage only when the model's forward applies prefix_seg_info by @danielhanchen in #17058 Studio: skip QKV fusion when projections are kept packed by @danielhanchen in #17132 Studio: streamed torchao weights no longer crash after a compile (Cannot swap t1 ... weakref) by @danielhanchen in #17141 MoE DDP negative control: accept the idle rank's missing gradient, whichever rank fails first by @danielhanchen in #17151 Studio Qwen-Image-2.1: reuse an unchanged condition image's VAE latents across renders by @danielhanchen in #17092 Studio: no strip behind Back to Hub over a background image by @shimmyshimmer in #17152 Studio: keep fp32 timestep and RoPE tables in fp32 on single-file loads by @danielhanchen in #17142 Studio LTX-2.5: one denoiser lease per CFG request, upload the next streamed encoder layer while this one runs by @danielhanchen in #17123 Studio SSM kernels: read the loaded torch build instead of a torch-import child per load by @danielhanchen in #16996 unsloth chat: parse the streamed Markdown when a frame is drawn, not on every chunk by @danielhanchen in #16994 Studio cached models: keep parsed README front matter per unchanged file by @danielhanchen in #17034 Studio VLM conversion: fetch repo image files on a pool, settle rows in order by @danielhanchen in #17029 Deep Research: keep a clean attachment inventory clean when inserting the terminal fallback by @danielhanchen in #16997 Studio diffusion runs: keep a passing full-load verdict per unchanged bundle for the run listing by @danielhanchen in #17033 Studio: load the chat runtime on use instead of in the startup bundle by @danielhanchen in #16999 Studio /v1/embeddings: count tokens in one batch and serialize float rows with tolist by @danielhanchen in #17030 Studio downloads: skip the GGUF requirement lookup on scoped progress polls by @danielhanchen in #17028 Studio chat: skip the transcript read when no message recorded a context fit by @danielhanchen in #16995 Studio tool output: strip ANSI in plain runs and find the tail without splitting the whole text by @danielhanchen in #16998 Studio llama.cpp: remember the positive reasoning budget probe per binary by @danielhanchen in #17031 Unsloth Studio / Desktop: queue image generations from the Images page and the API by @LeoBorcherding in #15005 Unsloth Studio: refuse Cosmos3 on diffusers 0.39 and 0.40 instead of failing at load or generate by @LeoBorcherding in #16745 Studio: each model's compile bundle holds only its own artifacts by @danielhanchen in #17110 Studio RAG: clip re-embedded chunks and embeddings API inputs with batched token counts by @danielhanchen in #17032 Studio: packed nvfp4 / mxfp8 layers follow a module dtype cast by @danielhanchen in #17131 Studio: run FLUX.3 in float32 on fp16-only GPUs (T4 renders were black) by @danielhanchen in #17124 Studio: bit-identical one-kernel RoPE for FLUX.3 eager attention (L4: -12% s/step) by @danielhanchen in #17136 Studio CLI tests: keep the --secure re-exec tests off the cloudflared download by @danielhanchen in #17129 RoPE: rotate Q and K in one launch each way under torch.compile by @danielhanchen in #17109 Studio: let peft 0.19 attach a LoRA to torchao int8 / fp8 weights of quantised single files by @danielhanchen in #17087 Studio MLX convert: pack 3 / 5 / 6-bit codes from Fortran-order or strided arrays by @danielhanchen in #17082 Studio diffusion: expandable VAE decode keeps a runtime allocator opt-in and restores the exact settings by @danielhanchen in #17076 Studio GGUF: load IQ1_S / IQ1_M / IQ2_XXS / IQ2_XS / IQ2_S / IQ3_XXS / IQ3_S on the diffusers path by @danielhanchen in #17071 Studio: re-merge only the layers of the LoRA adapters whose weight changed by @danielhanchen in #17102 Studio MiniMax-H3: normalise a reference video one VAE chunk at a time by @danielhanchen in #17096 Studio LTX-2.5: keep a fixed negative prompt in the encoder cache (refresh entries on a hit) by @danielhanchen in #17138 Studio diffusion: FLUX.3 modulation heads project once per sample; offloaded Base edits reuse the conditioning streams by @danielhanchen in #17059 Studio LTX VAE: keep only the open tile band on the card during tiled decode by @danielhanchen in #17060 Studio text encoders: keep the fp8 storage across a layerwise-cast forward instead of re-casting it by @danielhanchen in #17103 Context parallel: Ulysses for DeepSeek (V2 / V3 MLA, V3.2 indexer, V4 sink attention) by @MegaGimen in #15882 Studio: per-tool MCP switches at the end of Settings > Agents by @Etherll in #15629 Studio diffusion: keep single-file int8 weights as int8 on T4 / pre-Ampere NVIDIA and Intel XPU under offload by @danielhanchen in #17079 Studio: a denoiser-only LoRA change keeps the cached text embeddings by @danielhanchen in #17105 Studio: run the int8 / fp8 Triton widen on the operand's GPU by @danielhanchen in #17133 Studio MiniMax-H3: keep only the conditioner hidden state H3 reads (about 2 GiB less at 4096 tokens) by @danielhanchen in #17095 Studio: keep the LoRA fuse host backup off integrated GPU pins and inside the host RAM reserve by @danielhanchen in #17086 Studio video: fall back to the tiled VAE decode when the Linux / macOS CPU allocator refuses by @danielhanchen in #17075 Studio: queued image generations wait on their own pool, not the default executor by @danielhanchen in #17108 Studio: judge the compiled ROCm W8A8 linear per card, not once per process by @danielhanchen in #17078 Studio: let chat call Unsloth's own MCP tools by @Etherll in #15631 Studio diffusion: pair Qwen-Image-2.1's guidance branches per streamed block by @danielhanchen in #17111 Studio int8: fused MLPs and v2 Linears quantize their input with the probed one-pass kernel by @danielhanchen in #17114 Studio: the load prefetch skips chunks already in the page cache and budgets its read buffers by @danielhanchen in #17112 Studio image postprocess: convert to uint8 in row strips with one float32 copy each by @danielhanchen in #17091 Studio diffusion: launch the VAE and MOSS Triton kernels on the input's own card by @danielhanchen in #17099 Studio NVFP4: the protected branch compiles once from a cold decode table and accepts fp16 inputs by @danielhanchen in #17081 Studio upscaler: normalise seams and finalise uint8 in place on the host result by @danielhanchen in #17083 Studio GGUF: cast K-quant and legacy weights to bf16 in their own buffer by @danielhanchen in #17080 Studio MiniMax-H3 VAE: stitch batched decoder tiles into the frame as they arrive by @danielhanchen in #17097 Studio Sol-H3 VAE: decide window masks from the geometry, build each RoPE table once per decode by @danielhanchen in #17100 Studio int8 GEMV: wide scale groups tile inside shared memory, a grouped miss no longer demotes per-row Linears by @danielhanchen in #17106 Studio int8 GEMM: the unlisted-card race switch acts between loads by @danielhanchen in #17084 Studio FLUX.3: keep bf16 weights as stored on fp16-only GPUs and compute in float32 per layer by @danielhanchen in #17173 FSDP2 save: compose the writer across equal meshes, not only identical ones by @danielhanchen in #17158 Studio: run INT8 on T4 and other Turing GPUs instead of refusing it by @danielhanchen in #17167 Studio Qwen-Image-2.1: encode prompts without the text encoder's unused K/V cache by @danielhanchen in #17118 Studio Wan fused blocks: one Triton binary per kernel for every sequence length by @danielhanchen in #17101 Studio: fp8 denoisers on sm_89 no longer hang on the first render (CUDA graph replay under torch < 2.12) by @danielhanchen in #17166 Context parallel: Ulysses for Gemma 4 text models by @MegaGimen in #15883 Studio: install FlashInfer's prebuilt kernels into the vLLM engine env by @danielhanchen in #17063 Studio llama.cpp: the token-count memo renders again on a new day by @danielhanchen in #17090 Context parallel: Ulysses evidence for gpt-oss (attention sinks, MoE, aux off) by @danielhanchen in #17183 Studio video: log an automatic text-encoder pick after the load retries its hosted encoder by @danielhanchen in #17093 Studio tests: make the compiled-wrapper onload test hold its own weakref (follow-up to #17141) by @danielhanchen in #17189 Studio tests: the two-card compiled W8A8 test judges each card on its own arch (follow-up to #17078) by @danielhanchen in #17195 Context parallel: patch DPO / GRPO even when context_parallel is imported before the models by @danielhanchen in #17185 Studio diffusion: reload a released text encoder from its own safetensors files, no snapshot write by @danielhanchen in #17184 Studio: no header strips on macOS glass or over a plain background image by @shimmyshimmer in #17154 Backend CI: split the g-r pytest shard into g-l and m-r by @danielhanchen in #17187 Studio: read the hosted LTX-2.5 int8 checkpoint when the load's own plan offloads the denoiser by @danielhanchen in #17143 Backend CI: back off for about five minutes before giving up on the ruff-unsloth wheel by @danielhanchen in #17211 Studio MiniMax-H3: run the text stream once per render, pack the sequence in place, look up the streamed conditioner's vocabulary rows on the host by @danielhanchen in #17203 Studio diffusion: read the XPU's float64 support from the device (Arc A-series LTX RoPE) by @danielhanchen in #17070 Studio MiniMax-H3 Sol VAE: free expired Q/K/V slots before building the next, at most two live by @danielhanchen in #17085 Frontend CI: back off for minutes before giving up on the Bun download by @danielhanchen in #17221 Studio GGUF: dequantise bit-identical to llama.cpp, rounded once to the compute dtype by @danielhanchen in #17204 Studio: int8 ConvRot by default for FLUX.2 klein and FLUX.3 klein (LPIPS 0.10 to 0.04) by @danielhanchen in #17209 Studio diffusion: size CPU / MPS VAE decode tiles from free host memory (CPU 512 px decode 2.1x faster) by @danielhanchen in #17186 Studio: decode the FLUX.2 / FLUX.3 VAE in row strips (36-43% less decode VRAM, same speed) by @danielhanchen in #17202 Studio Qwen-Image-2.1: one GEMM for the dense gate + value and Q / K / V projections by @danielhanchen in #17214 Compiled RoPE: keep per-tensor launches when only Q or K needs a gradient by @danielhanchen in #17205 Studio MiniMax-H3: keep the whole denoiser forward compilable with the render text cache (follow-up to #17203) by @danielhanchen in #17226 Studio: keep segment CUDA graphs on unprotected NVFP4 steps by @danielhanchen in #17216 Studio: run Detailer inpaints on the image-generation pool by @danielhanchen in #17188 Studio LTX-2.5: auto text-encoder precision loads the hosted int8 ConvRot encoder by @danielhanchen in #17179 Studio MiniMax-H3: keep the conditioner read capture and tile decode traceable under torch.compile (#17095, #17097) by @danielhanchen in #17190 Studio compile cache: torch 2.7 keeps reused graphs in saved bundles (follow-up to #17110) by @danielhanchen in #17196 Studio diffusion: keep a streamed denoiser's blocks resident for the denoise phase, not the whole render (A100 8 GiB Qwen-Image-2.1 -18%, FLUX.3 -21 to -24%) by @danielhanchen in #17207 Studio FLUX.3 / Qwen-Image-2.1: prompt encodes keep only the hidden states they read by @danielhanchen in #17182 Studio CUDA graphs: the sm_89 fp8 stream-K guard reads every split card, not only the primary (follow-up to #17166) by @danielhanchen in #17194 Studio: install FlashInfer 0.7.3 prebuilt kernels for diffusion NVFP4 by @danielhanchen in #17215 Studio sd.cpp: pin the prebuilt resynced with upstream master-1050 by @danielhanchen in #17148 Studio: export chats as HTML and PDF by @shimmyshimmer in #17227 Studio diffusion: keep Qwen-Image-2.1 CFG pairing through its first-render compile by @danielhanchen in #17177 Studio LoRA fuse: charge the host RAM guard what the pinned backup really holds by @danielhanchen in #17191 Studio FLUX.3 manual cast: keep LoRA adapters at the fp32 compute dtype and out of the bf16 fold by @danielhanchen in #17201 Studio: give each rotated native int8 Linear its own ConvRot buffer (FLUX.3 INT8 on T4 failed its first render) by @danielhanchen in #17217 Studio FLUX.3: run attention in float16 on fp16-only GPUs (T4: -15% per image GGUF, -34% INT8) by @danielhanchen in #17223 Studio FLUX.3 on fp16-only GPUs: an explicit INT8 / FP8 quant reads the bf16 checkpoint as stored instead of widening the transformer to float32 by @danielhanchen in #17228 Studio FLUX.3: price float32 compute on fp16-only GPUs from the measured activation table (x2), so a GGUF Q4 denoiser stays resident on a T4 by @danielhanchen in #17219 Studio FLUX.3: keep reference-image modulation at one row per span in the joint blocks by @danielhanchen in #17181 Studio Qwen-Image-2.1: encode a condition image again after load_state_dict(assign=True) on its VAE by @danielhanchen in #17197 Tool activity gate: wait for each card to open before reading it, not a fixed sleep by @danielhanchen in #17234 Studio: a streamed GGUF denoiser's top-level weights re-point to their host copy instead of copying back every step (FLUX.3 Q8_0 T4: -11% per image) by @danielhanchen in #17222 Studio small-host INT8: armed layers compile once (eager-only call counters, follow-up to #17167) by @danielhanchen in #17193 RoPE: no recompile per sequence length when a compiled backward hands over a seq-major gradient by @danielhanchen in #17212 Studio FLUX.3: pass cuDNN attention the projections' strided Q/K/V instead of contiguous copies by @danielhanchen in #17180 Studio diffusion: keep Qwen-Image-2.1 scoped LoRA prompt reuse after the no-K/V wrapper, and key callable helpers fully by @danielhanchen in #17192 Studio: drop the live Glass blur while anything moves, on software-composited pages by @danielhanchen in #17007 Studio: stop the autoscroll follow flag restyling the whole thread in WebKit by @danielhanchen in #17035 Studio: open Settings and the shortcuts list with one panel render and one focus move by @danielhanchen in #17042 Studio: hold mid-scroll code block work on WebKit, fixing the #16865 scroll regression by @danielhanchen in #17149 Studio: run FP8 on T4 and other Turing GPUs as weight-only fp8 instead of refusing it by @danielhanchen in #17218 Studio: stop the Windows Terminal from creating nul files through Git Bash by @danielhanchen in #17010 Studio: save Image LoRA training settings as presets by @danielhanchen in #17019 Studio RAG: pause linked-folder indexing while a local chat generates by @danielhanchen in #17018 Studio LTX-2.5: compact conditioning transport and stage-two connector reuse by @danielhanchen in #17232 Studio: load INT8 instead of NVFP4 on pre-Blackwell GPUs by @danielhanchen in #17229 Studio: keep single-file int8 denoiser weights as int8 on CPU instead of expanding them to float32 by @danielhanchen in #17178 Studio: fuse the int8 ConvRot rotation into its producers on FLUX.2 / FLUX.3 (ConvRot cost +8.7% to +0.5%) by @danielhanchen in #17230 Studio diffusion: price a streamed Qwen-Image-2.1 edit's reference from its token count, not the flat allowance (A100 8 GiB edit -15%, ahead of the baseline) by @danielhanchen in #17233 Studio chat: only a note followed by its wrapper starts an attachment by @YakshithK in #16915 Studio LTX-2.5: use two CUDA cards, the sidecar text encoder priced on its own card by @danielhanchen in #17077 Condense comments added since the last cleanup: Studio frontend and Tauri by @danielhanchen in #16109 Condense comments added since the last cleanup: installers, scripts and CI by @danielhanchen in #16112 Studio update: stop size-checking the managed Windows launcher by @debojyoti-dev in #16901 Studio MiniMax-H3: keep the released bf16 denoiser resident beside a streamed conditioner on ~100 GB cards by @danielhanchen in #16877 Add configurable voice context to reduce Qwen3-ASR memory by @Imagineer99 in #15810 Fix file drops in new project chats by @Imagineer99 in #14472 Studio video: run LTX-2.5 FP8 on a 24 GB card by streaming the hosted fp8 DiT instead of refusing it by @danielhanchen in #17238 Studio: pause Glass blur through bursts of keys and presses under software compositing by @danielhanchen in #17237 Condense comments added since the last cleanup: Studio backend by @danielhanchen in #16110 Condense comments added since the last cleanup: core, CLI and tests by @danielhanchen in #16111 [Feature] Unsloth Studio / Desktop: let the mic send audio straight to any audio-input model by @LeoBorcherding in #16976 Studio: bake the blurred chat wallpaper once instead of re-running the blur on every scroll frame by @danielhanchen in #17253 gloo-tp tests: a rank that finished is not failed by an abort in interpreter teardown by @danielhanchen in #17244 Studio ROCm layer split: probe secondary cards read-only, at the attention dtype by @danielhanchen in #14655 Studio ROCm VAE attention: math SDPA in the VAE's bf16/fp16 dtype, not an fp32 upcast by @danielhanchen in #17252 Studio CFG-parallel: treat an NVFP4 precision-state flip as a cold call so the two branches never compile at once by @danielhanchen in #14854 Studio: load image / video / text-encoder weights from .safetensors first, reuse a cached .pt / .bin when its sha256 matches the Hub by @danielhanchen in #17061 Studio tests: stub stt-context when mic-target loads the voice settings store by @danielhanchen in #17293 Make shared composer drafts opt-in by @Imagineer99 in #16548 Studio: give FLUX.3's per-mode block stacks their own offload groups (5 GiB L4 int8: -17.6% per image) by @danielhanchen in #17122 Studio: serve the hosted Wan2.2-A14B int8 pair to an offloaded load as torchao-free W8A8 layers by @danielhanchen in #15240 Studio diffusion: auto attention keeps native flash SDPA on A100 / L4 class cards by @danielhanchen in #17290 Studio tests: pin the H3 cached-prequant lookup offline and cover the online twin rule by @danielhanchen in #17309 Studio diffusion: split the batch across two cards for LTX-2, Chroma, Qwen-Image, HunyuanImage and Krea-2, and pair Cosmos 3 guidance by @danielhanchen in #17278 Studio tests: skip the FLUX.3 group-offload test where diffusers is not installed by @danielhanchen in #17317 Studio ROCm: use the measured fastest GEMM solution table for diffusion denoisers on gfx1151 by @danielhanchen in #17299 Long-chat gate: settle pointer targets before a trusted hover or click by @danielhanchen in #17274 Long-chat gate: measure streaming a reply into the chat by @danielhanchen in #17269 Long-chat gate: measure every appearance setting or allowlist it with a reason by @danielhanchen in #17262 Studio native diffusion: run the text encoders on a second CUDA card when one card cannot hold the load by @danielhanchen in #17276 Studio diffusion: keep the fused AdaLN census out of compiled FLUX.1 blocks (recompile limit on first render) by @danielhanchen in #17310 Use one glass surface for the docked Windows browser by @Imagineer99 in #16754 Studio: title tooltips stop walking the whole thread on every pointer move by @danielhanchen in #17304 Studio: Glass Settings pane scrolls on the compositor; API monitor fade pauses its blur (software compositing) by @danielhanchen in #17308 Studio: keep the view_image picture visible when its tool card is collapsed by @keyiyu0331 in #17174 Studio: stream a reply without re-rendering every message client per token by @danielhanchen in #17305 Studio: browser pane opens without per-frame thread work on long chats by @danielhanchen in #17282 Windows install: keep typer below 0.29 so studio setup does not hold pydantic_core open by @danielhanchen in #17319 Studio ROCm TunableOp: label the results file with the HIP build string, not a raw read by @danielhanchen in #17325 Studio: colour chat code blocks from the worker while scrolling and streaming by @danielhanchen in #17160 Studio: key group hover / focus-within / active on the group class (WebKit long-chat menus and dialogs) by @danielhanchen in #17323 Studio: modal opens stop re-laying out long chats in WebKit by @danielhanchen in #17284 Studio: WebKitGTK Glass drops the sidebar's live blur and Frosted's grain tile (long-chat scroll) by @danielhanchen in #17327 Unsloth Studio: restyle a video with Cosmos3-Nano-Sim-Transfer on sd-server by @LeoBorcherding in #16975 Fix Windows wallpaper titlebar and Transparent glass surfaces by @Imagineer99 in #16959 Studio CI: WebKit leg for the long-chat gate, report-only paint / CPU length check, recheck gap fixed by @danielhanchen in #17331 Studio ROCm RDNA int8: Triton W8A8 kernels and fused Qwen-Image-2.1 projections (gfx1151 INT8 -12% per step, explicit int8 -26% to -34%) by @danielhanchen in #17334 Studio: open the browser pane without React's Suspense reveal hold by @danielhanchen in #17335 Long-chat gate: block on CPU / paint growth with chat length, judged on fresh-page rechecks by @danielhanchen in #17336 Studio: share the FLUX.3 joint block int8 ConvRot quant under the span forward by @danielhanchen in #17337 Studio: reroute only official NVFP4 checkpoints to INT8 on pre-Blackwell GPUs by @danielhanchen in #17343 Studio FLUX.3: probe strided cuDNN attention on every layer split card by @danielhanchen in #17341 Studio diffusion: do not let functools.wraps copy the prompt-cache transparent-wrapper marker by @danielhanchen in #17348 Studio sd.cpp: only block-split on builds whose --auto-fit takes on|off by @danielhanchen in #17346 Studio ROCm layer split: apply the common SDPA guard only once the split fits by @danielhanchen in #17347 Studio: an unverified cached .pt no longer demotes the baseline-format safetensors twin by @danielhanchen in #17342 Studio sd.cpp: send MiniMax-H3 reference video and audio to the resident sd-server by @danielhanchen in #17349 Studio: glass toast surface for the runtime, Studio and Desktop update cards by @shimmyshimmer in #17352 Studio: restore ROCm TunableOp state when the last diffusion pipeline unloads by @danielhanchen in #17340 Studio: compute ROCm VAE half-math SDPA per call instead of flipping the global flag by @danielhanchen in #17350 Studio sd.cpp: retry an image render that sd-server refused for VRAM under half the budget by @oobabooga in #17339 Studio: open a long chat at its latest message on Chromium by @danielhanchen in #17351 Studio: open and close dialogs and menus in long chats without a full thread re-layout on Firefox by @danielhanchen in #17357 Studio: fit a GGUF reload only after the stopped llama-server's VRAM shows as free by @oobabooga in #17345 Studio MiniMax-H3: skip the up-front denoiser slab pin when every block will be resident (-16 GiB host peak on A100 40 GB) by @danielhanchen in #17354 Studio MiniMax-H3: keep conditioner layers on the card beside a resident denoiser, size its host pins to their weights by @danielhanchen in #17355 Studio MiniMax-H3: park the VAEs when a render returns, so later requests keep their residency budget by @danielhanchen in #17356 Studio diffusion: keep no pinned host copy of a resident Qwen-Image-2.1 text encoder (L4 host RAM -8.1 GiB, peak RSS 25.3 -> 16.6 GiB) by @danielhanchen in #17359 Studio sd.cpp: send the simple scheduler so flow families get the diffusers sigma grid by @danielhanchen in #17360 Studio diffusion: keep a hosted checkpoint's header across processes, so an 8 GiB Qwen-Image-2.1 load skips its FP8 header read (L4 load -1.8 s) by @danielhanchen in #17353 Long-chat gate: Firefox leg for dialogs, menus, browser pane and typing by @danielhanchen in #17361 Unsloth Studio: Cosmos3 256p, NVIDIA's frame ranges and a per-model frame-rate picker by @LeoBorcherding in #16737 Decode CUDA graphs on by default for every FastModel model (2-8x on T4, tokens equal to eager) by @danielhanchen in #16587 Studio diffusion: price a Qwen-Image-2.1 edit's reference on the group tier so 16 GiB edits stop hooking the resident transformer (L4 first edit -19%, -8.4 GiB host RAM) by @danielhanchen in #17364 Studio: reuse Windows ROCm Get-Counter readings for display polls like amd-smi by @danielhanchen in #17306 unsloth CLI: check a Studio pid with psutil before spawning tasklist on Windows by @danielhanchen in #17271 Studio Images X/Y plot: shrink grid cells off the event loop by @danielhanchen in #17265 Studio debug logs: list log names instead of building a Path per file on each poll by @danielhanchen in #17302 Studio library: cache Intl formatters for per-row dates and sizes by @danielhanchen in #17251 Studio research: run each step's knowledge base and MCP lookups alongside its web search by @danielhanchen in #17287 H3 rotation park test: read video.py as UTF-8 by @danielhanchen in #17370 Unsloth Studio: Cosmos3-Nano video continuation (extend a clip from its last or first frames) by @LeoBorcherding in #16741 Studio: keep the browser feature index out of startup by @danielhanchen in #17280 Studio: reduce Liquid Glass paint cost when WebKit paints on the CPU by @oobabooga in #16960 Studio embeddings: parse the GGUF body only when read and once, and serialize the Studio embedder payload off the event loop by @danielhanchen in #17286 Studio dataset preview: hand a JSON file that is not JSON Lines to the loader without reading it whole by @danielhanchen in #17277 Studio chat search: skip metadata_json when building the search index batch by @danielhanchen in #17249 Studio GGUF variants: reuse the request's Hub listing for the MLX fallback by @danielhanchen in #17279 Studio: skip superseded wallpaper writes and no-op history / bookmark persists by @danielhanchen in #17298 Studio Windows tools: stop the trusted git lookup at the first trusted candidate by @danielhanchen in #17270 Studio: memoize the nonce-free Content-Security-Policy on its inputs by @danielhanchen in #17246 Studio: keep marked out of startup by moving the export format names to their own module by @danielhanchen in #17250 Studio Data Recipes: keep unstructured seed upload file IO off the event loop by @danielhanchen in #17266 Unsloth Studio: list Cosmos3-Nano in the Video picker and route it from the Hub by @LeoBorcherding in #16743 Studio dataset detection: read only the still-missing columns when filling a sparse first row by @danielhanchen in #17248 Studio: pause display-only training polls while the tab is hidden by @danielhanchen in #17294 Studio: import HfApi on first use instead of at backend start by @danielhanchen in #17273 unsloth studio stop: return as soon as the server exits instead of on the next half-second step by @danielhanchen in #17283 Patch sentence_transformers when it is first imported instead of importing it with unsloth by @danielhanchen in #17267 Studio Metal: read the Apple wired headroom once per GGUF placement by @danielhanchen in #17275 Studio chat: skip the display strip chain for history text with no tool markup by @danielhanchen in #17264 Studio RAG: delete a retired folder's or scope's chunks in batches instead of one index scan per document by @danielhanchen in #17281 Studio: cache the native-template markup profile per tokenizer in renderable_tool_catalog by @danielhanchen in #17288 Studio Anima LoRA: rule out kohya splits that name no registered module before probing by @danielhanchen in #17247 Diffusion guard test: match the calibrated placement and ControlNet kwargs structurally by @danielhanchen in #17376 Studio: resolve the nested Radix primitive / slot / context copies to one version by @danielhanchen in #17301 Studio: keep speaker notes when a PowerPoint file is attached in chat by @NilayYadav in #17258 Studio: keep searching the web past a lone Wikipedia result by @NilayYadav in #17242 Studio Windows GPU inventory: run the PowerShell adapter scan only when AMD or Intel records need it by @danielhanchen in #17272 Studio installers: ask for the latest llama.cpp / whisper.cpp release tag once per run by @danielhanchen in #17289 Studio: read once per tab return instead of on both visibility and focus by @danielhanchen in #17300 Studio: keep the Ask before downloading programs setting across browsers by @NilayYadav in #17240 Studio: turn Reduce motion on by default by @shimmyshimmer in #17377 Skip the duplicate tokenizer load on transformers 5 when it can only copy the fast tokenizer by @danielhanchen in #17318 Studio: keep a user-set Hermes api_key in unsloth start hermes by @NilayYadav in #17320 Studio: train Phi-4-mini on its own chat template by @NilayYadav in #17312 Studio MiniMax-H3: run one denoiser forward per requested step (20 steps ran 19) by @danielhanchen in #17362 Slot-clamp lint: accept the allow marker on any line of a wrapped assignment by @danielhanchen in #17384 Fix CPU pipeline evaluation loss normalization by @0xalkinun in #17291 Studio: keep hidden spreadsheet rows in chat attachments by @NilayYadav in #17314 Studio: save LoRA and merged exports of a run to separate folders by @NilayYadav in #17260 Decode CUDA graphs for Qwen3.5 / Qwen3.6 hybrid linear-attention caches by @danielhanchen in #17324 Studio: stop audio models learning to answer None for blank transcripts by @NilayYadav in #17257 Studio: hold the login rate limit on the Colab public link by @NilayYadav in #17259 Studio: end plain-text training rows with EOS when Format is Auto by @NilayYadav in #17255 Decision models on MLX: refuse a config.json model_file unless trusted by @danielhanchen in #17391 AsyncGRPO: require a per-run secret on the local vLLM server; validate BF16 AdamW state on load by @danielhanchen in #17390 Studio: cap request bodies on every route and decode zstd bodies one at a time by @danielhanchen in #17388 Studio: do not start a queued fine-tuning job for a disabled or deleted account by @danielhanchen in #17392 Studio: edit_file and Rewind write through an open chat-folder descriptor on POSIX by @danielhanchen in #17389 Studio: open update-notes links in the Desktop app, treat .term and .tcsh downloads as code by @danielhanchen in #17394 Long-chat gate coverage: measure forced motion now that Reduce motion is on by default by @danielhanchen in #17395 Studio: make keyword search find Chinese, Japanese and Korean words by @NilayYadav in #17261 Studio video: a stored peak no longer widens a new process's first render; an out-of-memory render retries streamed (LTX-2.5 16 GiB returning home) by @danielhanchen in #17373 Studio: harden diffusion adapters, quant scales, prompt LoRA tags, Hub clips, short audio and restyle uploads by @danielhanchen in #17396 Settings sync: save the shared composer draft preference with the account by @danielhanchen in #17399 Notebook GGUF export: keep the converter to the load's trust_remote_code decision by @danielhanchen in #17397 Bump trufflesecurity/trufflehog in / by @dependabot[bot] in #17243 Studio: centre dialogs on whole pixels so their text stays sharp by @wasimysaid in #17383 Studio: fix embedding fine-tunes failing at the first step by @NilayYadav in #17311 Studio: make MLX and safetensors models call a forced tool by @NilayYadav in #17241 Studio: say when a file added while creating a project fails to index by @NilayYadav in #17313 Studio: open the managed account folder browser on symlinked installs by @NilayYadav in #17321 Studio: remove Liquid Glass stripes and greeting rings, and the black browser tab strip on Linux by @wasimysaid in #17380 Studio: keep every frame when downloading a video as WebM by @NilayYadav in #17239 Studio GGUF vision chat: reuse the image normalization for history images sent again by @danielhanchen in #17263 Studio: ask before Python shells out through asyncio or pty by @NilayYadav in #17254 Studio: send video to custom OpenAI-compatible connections by @NilayYadav in #17322 Studio: restore real Glass on the Linux app (live blur, fills, strength and the Windows-style title bar) by @wasimysaid in #17387 Studio: keep deferred MCP tools in place when a listed tool runs again by @oobabooga in #17326 [CS146S Stanford] Studio web fetch: a bare & stays as written, and escaped code stays in its code block by @mrityu75 in #17200 Studio: stop image loads exhausting Windows commit on Intel Arc iGPUs by @danielhanchen in #17378 Opt-in growable KV cache for generic decode graphs by @danielhanchen in #17296 GGUF export: keep remote_code_approved out of the save_pretrained kwargs by @danielhanchen in #17404 docker(rocm): run gfx103X cards as gfx1030 when the image's torch lacks them by @hMelchisedek in #17150 fix(studio): align cached GGUF picker size by @Aafimalek in #17125 Remove unreferenced private helpers, constants and an if False block by @danielhanchen in #16569 Add retry button to download panel for failed/cancelled downloads (CS 146S) by @vijkomal in #17225 fix(studio): support explicitly selected virtual bridge LAN addresses by @alanhuang05 in #17165 Studio diffusion: load Qwen-Image-2.1's hosted int8 text encoder on float16 cards (T4 Turbo INT8 82.7 -> 31.9 s/image) by @danielhanchen in #17379 Studio LTX-2: stream the text connectors under streaming offload so the denoise keeps more transformer groups resident (A100 16 GiB -16% per clip) by @danielhanchen in #17386 Studio: set scrollbar-color in rgb so Reduce motion stops restarting a transition on every element by @danielhanchen in #17413 Studio: leave room for thinking in a Custom connection's chat title request (#17332) by @indrajeetapache in #17363 Installers: require unsloth-zoo>=2026.11.4 to match the pyproject floor by @danielhanchen in #17414 Installer: clear damaged uv cache entries before retrying an install by @danielhanchen in #17398 Studio: mount the API monitor inside a host that stays mounted, so opening it in Glass stops restyling the app by @danielhanchen in #17418 Studio FLUX.3: keep float16 on fp16-only GPUs behind a residual rescale (T4 INT8 -24% per image, GGUF Q8_0 -73%) by @danielhanchen in #17385 Scan baseline: re-review the seven unsloth-zoo sites 2026.11.4 moved by @danielhanchen in #17421 Studio: keep more of a streamed FLUX.3 klein int8 transformer resident at small budgets (L4 6 GiB -4.2% per image, H2D 56.9 -> 43.9 GiB) by @danielhanchen in #17419 Studio MiniMax-H3: feed the SwiGLU halves straight into the down projection's int8 activation quant (L4 -0.85% s/step, bit-identical) by @danielhanchen in #17422 Update idempotency: let the offline test survive a version bump that is not on PyPI yet by @danielhanchen in #17428 Studio: ask before running Python fed to Terminal through a heredoc by @NilayYadav in #17256 Studio tests: read video.py as UTF-8 in the H3 fused SwiGLU test by @danielhanchen in #17437 Unsloth Studio: show which models have saved run settings and reset them from the model menu by @LeoBorcherding in #12608 Studio Terminal: give the heredoc quote scanner its own substitution helper by @danielhanchen in #17442 Studio sd.cpp: pin the prebuilt resynced with upstream master-1056 by @danielhanchen in #17459 Studio: stop Reduce motion turning on a transition for every element by @danielhanchen in #17466 Studio NVFP4: tune mm_fp4 once per M bucket and keep the tactics on disk by @danielhanchen in #17467 Studio: neutral NVFP4 gate field names and refreshed checkpoint hashes by @danielhanchen in #17469 Studio NVFP4: 4over6 block scales at load and in the prequant builder by @danielhanchen in #17468 Tests: clear the cached tensor parallel layout before every test by @danielhanchen in #17472 Studio: MiniMax-H3-Image text to image and reference edits by @danielhanchen in #17453 Flex attention decode steps for Gemma 4 and Qwen3.5 / Qwen3.6 under decode CUDA graphs by @danielhanchen in #17329 Studio tests: give the llama-server owner tests a wait cap a loaded runner can meet by @danielhanchen in #17488 Studio: keep chat text out of the header with Glass or a background by @NilayYadav in #17429 Studio: ask before Python runs a shell command passed by reference by @NilayYadav in #17456 Studio: let MCP run_recipe set a row count by @NilayYadav in #17427 Studio: stop web results crowding attached files out of Deep Research by @NilayYadav in #17434 Studio: wrap long unbroken text in PDF chat exports by @NilayYadav in #17425 Studio: make Gemma 4 call a forced tool by @NilayYadav in #17432 Studio: keep the seeded admin password out of pages relayed by a proxy by @NilayYadav in #17454 Studio: stop showing a stray after an OpenAI model calls a tool by @NilayYadav in #17455 Studio: fit max_tokens to the window on vLLM and SGLang tool requests by @NilayYadav in #17426 Studio: keep the LoRA on in Chat after a Base vs LoRA compare by @NilayYadav in #17433 Studio: let managed accounts use gated models with their own HF token by @NilayYadav in #17172 Studio: keep built-in tools on when an MLX chat has several images by @NilayYadav in #17423 Studio: compact long chats on an Ollama connection to its loaded window by @NilayYadav in #17457 Studio: reconnect a chat reply whose connection died silently by @NilayYadav in #17416 Studio: always show loaded GGUF weight placement by @oobabooga in #17449 Studio MiniMax-H3: opt-in 4-bit GGUF text encoder for hosts that cannot keep int8 resident by @danielhanchen in #17479 Studio: edit images through an external connection's /images/edits endpoint by @oobabooga in #17448 Gemma 3 generate test: find the non-flex scope by call, not by its exact text by @danielhanchen in #17497 Studio sd.cpp: pin the prebuilt with the aarch64 K-quant GEMM and macOS Accelerate routing by @danielhanchen in #17489 TP / FSDP2 saves: pass results between ranks as JSON, not pickle by @danielhanchen in #15510 Studio: accept chat videos up to 256 MB by @danielhanchen in #17400 Context parallel: eval loss is the sequence's token mean, not cp_size times it by @lonexreb in #17049 Mask GRPO completions cut by the vLLM server's context limit by @danielhanchen in #16650 Pipeline parallel on the CPU backend: weight the MoE router aux loss as TRL's SFTTrainer does by @debojyoti-dev in #17072 Ulysses: dynamic and longrope RoPE see the whole sequence's positions by @lonexreb in #17051 Studio desktop: serve the web UI behind an allowlisted reverse proxy (fixes #17420) by @debojyoti-dev in #17435 RoPE: compile stock positions without full-context caches by @danielhanchen in #16182 Studio: train jobs targeting embed_tokens or lm_head alone, not batched by @danielhanchen in #16653 Multi-LoRA: keep the backward dH buffer in fp32 by @danielhanchen in #16647 Studio: MLX chats with a picture keep going when the context fills by @NilayYadav in #17424 Multi-LoRA: refuse 8-bit bitsandbytes base layers by @danielhanchen in #16648 Studio: no RAM fallback mark in the model picker when there is no GPU by @shimmyshimmer in #17516 Studio: pin MiniMax-H3-Image (Fast FP8) at the top of the image picker by @danielhanchen in #17502 New Contributors @MegaGimen made their first contribution in #15880 @0xalkinun made their first contribution in #17291 @hMelchisedek made their first contribution in #17150 @vijkomal made their first contribution in #17225 @alanhuang05 made their first contribution in #17165 Full Changelog: v0.1.916-beta...v0.1.917-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:
2026-11-12T03:07:16.792Z
Categories:

Related to Unsloth v0.1.917-beta:

Post a Comment/Report Broken Link: