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
Studio: queue /v1/videos while a video renders and batch waiting /v1/images/generations requests (#15613)
85744ff oobabooga committed 24 days ago
Tests: bring three new tests in line with the port and encoding guards (#15745)
ef726b9 Daniel Han committed 24 days ago
Studio: show cache sizes and clear individual Hugging Face model copies (#15488)
b489a92 Daniel Han committed 24 days ago
Studio diffusion: run the fused-dequant int8 GEMM on untuned sm80+ archs behind a probe and latency check (#15654)
36bd1af Daniel Han committed 24 days ago
Run xformers attention on the tensors' device (device_map across GPUs gave NaN loss) (#15712)
ab35854 Daniel Han committed 24 days ago
Keep an earlier model's smart offloading when a short sequence model loads (#15711)
010b49f Daniel Han committed 24 days ago
Studio: advertise the chat-title repair guards on the thread lists instead of fetching /openapi.json (#15685)
ae41442 Daniel Han committed 24 days ago
Studio: precompile bytecode after setup so the first launch does not (#15646)
1ba04b0 Daniel Han committed 24 days ago
Studio: compose the FP8 and NVFP4 compile cache tags in a fixed order (#15692)
d513620 Daniel Han committed 24 days ago
Studio: acquire decision datasets alongside model loading (#15481)
e7afc20 Daniel Han committed 24 days ago
Studio: report first-step training and gradient accumulation (#15480)
4ac4e91 Daniel Han committed 24 days ago
Studio: answer keyless admission on the event loop when nothing needs reading (#15683)
c79acda Daniel Han committed 24 days ago
Studio desktop: leave the startup screen on server-port instead of at the next 500 ms poll (#15673)
31665e7 Daniel Han committed 24 days ago
Studio: log a live generation rate while a reply is still streaming (#15390)
940a42a Nilay committed 24 days ago
Faster import unsloth and from_pretrained: fewer GC passes, skip regexes that cannot match (#15560)
89a8443 Daniel Han committed 24 days ago
Studio: walk each GGUF tokenizer array once per process and seek past it on later reads (#15687)
d2d6f76 Daniel Han committed 24 days ago
Studio desktop: share one loopback HTTP client per timeout (#15675)
e4d513f Daniel Han committed 24 days ago
Studio: make Stop work while an MLX model is still reading the prompt (#15615)
500f6db Nilay committed 24 days ago
Studio: count sd-cli WebM frames from VP8 packet headers in inspect_video (#15699)
badc8a0 Daniel Han committed 24 days ago
Studio: visible light-mode borders, aligned media pickers, new image icons, modern login (#15726)
bb807b7 Michael Han committed 24 days ago
Studio desktop: run -h and desktop-capabilities at once on a capability cache miss (#15672)
f8dad3f Daniel Han committed 24 days ago
Studio desktop: start the launch preflight's CLI -h probe during setup (#15671)
9594abc Daniel Han committed 24 days ago
Studio: stop re-stripping an open tool call body on every streamed token (#15707)
7945fb8 Daniel Han committed 24 days ago
Studio: mention ARM64 in Windows sandbox architecture guidance (#15494)
d28537f Daniel Han committed 24 days ago
Studio: look up model picker rows by index map instead of a scan per row (#15674)
8b0f6bd Daniel Han committed 24 days ago
Studio: delete a RAG document's index rows in one scan per table (#15661)
9791b20 Daniel Han committed 24 days ago
Studio: read each dist-info METADATA once per install check (#15690)
afd008e Daniel Han committed 24 days ago
Studio: render /openapi.json once per schema, off the event loop (#15684)
b0bfa3c Daniel Han committed 24 days ago
Studio: answer ASCII tool result budgets without a per-character scan (#15676)
c0b6758 Daniel Han committed 24 days ago