Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Commits on main
Gemma 4 packed SFT: efficient kernel for the 512-wide global layers, processor on text-only data (#13384)
37f9088 Daniel Han committed 1 month ago
Windows installer: always build Studio on a uv-managed Python, never one already installed (#13399)
e210476 Daniel Han committed 1 month ago
Pack Qwen3.5 / Qwen3-Next hybrids by default with document boundaries in the linear-attention kernels (#13383)
d5824d0 Daniel Han committed 1 month ago
Studio setup: install causal-conv1d and mamba-ssm prebuilt wheels next to flash-attn (#13396)
9f4a868 Daniel Han committed 1 month ago
Keep syntax colour on settled thinking fences in tilde, four-backtick, indented and stopped forms (#13415)
d6cfacf Daniel Han committed 1 month ago
Studio: keep Data Recipe runs whose structured fields change type between rows (#13407)
4e49365 Daniel Han committed 1 month ago
Studio: browser page zoom on desktop: zoom keys and clipping (#13401)
3075b73 Michael Han committed 1 month ago
Studio: keep the browser page below its toolbar while the split is dragged (#13400)
d6182bb Michael Han committed 1 month ago
FastLanguageModel.for_training: undo the UNSLOTH_RETURN_LOGITS=1 that generate() forces (#13392)
d4ea5a5 Daniel Han committed 1 month ago
Packed rows on sliding-window layers take varlen on the transformers path, plus opt-in FA4 / FA3 for packed training (#13326)
8399968 Daniel Han committed 1 month ago
Studio: document the intentional torch and torchcodec floor gap (#10635)
135dca3 Rudra Bhaskar committed 1 month ago
Opt-in DDP batched_grad_copy (UNSLOTH_DDP_BATCHED_GRAD_COPY=1) (#13350)
8892066 Daniel Han committed 1 month ago
Studio: opt-in MLX image engine for Apple Silicon (Z-Image) (#13379)
10bddef Daniel Han committed 1 month ago
Studio: refresh the context meter when a tool loop hands off to a tool (#13356)
1f1b9e7 Daniel Han committed 1 month ago
Studio: judge hosted pre-quant checkpoints before downloading them, and make the int8 fingerprint torchao-version-neutral (#13339)
2762603 Daniel Han committed 1 month ago
Studio MPS: opt-in fp16 compute with fixed headroom for Z-Image (#13375)
d3e21f3 Daniel Han committed 1 month ago
Studio: refuse LTX-2.5 instead of loading it as LTX-2 (#13370)
cfe6d93 Daniel Han committed 1 month ago
Studio: sd.cpp on Macs, CPU backend on a paravirtual GPU and no direct conv on Metal (#13369)
0464be9 Daniel Han committed 1 month ago
Sandbox delete test: only the chat's detached tree counts as the delete (#13403)
072c44c Daniel Han committed 1 month ago
Keep packed documents apart on Unsloth's own models (flash path, eval, cached eval) and match transformers' Mistral window (#13313)
bef1c3d Daniel Han committed 1 month ago
Studio: per-call output cap for python/terminal tools (superseded by #13228) (#10398)
4510e7f Shreehitha Arushan committed 1 month ago
Studio: let a project run its tools in an existing local folder (#13353)
15b2867 Daniel Han committed 1 month ago
Switch between training and inference mode in one tree walk (#13374)
0edcd85 Daniel Han committed 1 month ago
Studio: hide the response-details badge without a fade (#13397)
6060574 Michael Han committed 1 month ago
Honor padding masks in Llama and Qwen3 cached attention (#13256)
60a006d Arthur committed 1 month ago