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
GRPO: log the clip_ratio metrics for loss_type luspo (#17537)
eb34918 Vineeth Sai Varikuntla committed 8 days ago
Studio Desktop: bump zip, ctutils and rustix in Cargo.lock for security advisories (#17591)
6c4630a Daniel Han committed 8 days ago
Studio: refuse a Train Split Start past the last row instead of training on one row (#17303)
0b4860e Vineeth Sai Varikuntla committed 8 days ago
Studio: keep a chat's own settings when its send waits on a model load (#17565)
be23b57 oobabooga committed 8 days ago
Llama 3.x and GPT-OSS templates: treat tool_calls None as no tool call (#17176)
211f0ee Vineeth Sai Varikuntla committed 8 days ago
Images/Video: hide the pill's loaded state during eject and fix unload races (#17551)
b807d41 Lee Jackson committed 8 days ago
Studio frontend: bump node-forge to 1.4.3 and seroval to 1.7.5 (#16258)
dad5c73 dependabot[bot] committed 8 days ago
Strip Llama 3's bos_token set statement when deduping a chat template (#17316)
ad201d7 Vineeth Sai Varikuntla committed 8 days ago
Let the qwen3 chat template render tool-call turns with null content (#17175)
9fdbdd5 Vineeth Sai Varikuntla committed 8 days ago
Studio: read decision JSONL rows that hold U+2028, U+2029 or U+0085 (#17307)
4a7d860 Vineeth Sai Varikuntla committed 8 days ago
push_to_hub_gguf: put the user's tags in the README card (#17438)
e0ba54e Vineeth Sai Varikuntla committed 8 days ago
Qwen 2.5 template: use the chosen system message when tools are passed (#16932)
300409b Vineeth Sai Varikuntla committed 8 days ago
CLI: refuse unsloth train when num_epochs and max_steps are both 0 (#17116)
d26a6ce Vineeth Sai Varikuntla committed 8 days ago
Tests: read the titlebar-height mirror through fullscreen's customTop (#17592)
2e1f00c Daniel Han committed 8 days ago
Studio: preserve guidance for FLUX.1-dev single-file models (#17567)
973811b oobabooga committed 8 days ago
Studio: let an agent reply's round tail leave from the end of the list (#17501)
f2572a6 Daniel Han committed 8 days ago
Studio: persist local-checkpoint tier probe verdicts across workers (#17496)
82f21d1 Daniel Han committed 8 days ago
Honor the Vulkan opt-in for sd.cpp on Windows Intel GPUs (#17528)
a5522c0 Lee Jackson committed 8 days ago
Studio: Stop and eject interrupt a running image generation (#17559)
3b8286c Nilay committed 8 days ago
Bump github/codeql-action init and analyze to v4.39.3 (#17584)
61d752f dependabot[bot] committed 8 days ago
Studio: stop each chat switch from keeping the previous thread alive (#17498)
4e50fbc Daniel Han committed 8 days ago
get_model_name: keep the fetched upstream mapper for an hour instead of refetching it on every load (#17490)
20d9ab5 Daniel Han committed 8 days ago
Studio: memoize the RAG roster's cleaned document names (#17494)
e33cb8d Daniel Han committed 8 days ago
Studio: cache sd.cpp --help output per binary build (#17491)
192bce6 Daniel Han committed 8 days ago
Studio: fit long tool results to the window on connections (#17553)
fccc8c4 Nilay committed 8 days ago
Fix header gradient banding on plain glass backgrounds (#17403)
289d9a7 Lee Jackson committed 8 days ago
Studio: keep a Python script's output when a terminal call times out (#17555)
b5cf379 Nilay committed 8 days ago
Studio: size Deep Research prompts to the window a self-hosted connection serves (#17554)
32b328d Nilay committed 8 days ago