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
Turn the CUDA graph off for MiniMax-H3: measured 1.0018x for 3.93 GB (#11175)
7e08979 Daniel Han committed 2 months ago
Fall back to HTTPS when hf_xet is installed but cannot be imported (#11167)
6ed7f22 Daniel Han committed 2 months ago
Stop a removed torchao class from killing plain LoRA (#11168)
fff6ee6 Daniel Han committed 2 months ago
Studio: serve an installed MLX checkpoint whose config drops its vision sub-config (#10977)
b7d4a62 Long Yixing committed 2 months ago
Define everything Refresh-Environment calls before the PSModulePath test runs it (#11194)
dc151dd Daniel Han committed 2 months ago
Studio: keep a reasoning page start below its end (#10922)
3334e3c Daniel Han committed 2 months ago
Fix the cold unsloth.save import, the preview role map and the GPU memory cap (#11077)
e15a75e Daniel Han committed 2 months ago
Take the compile-cache bundle write off the image generation request path (#11009)
075a3cb Daniel Han committed 2 months ago
Diagnose torch and triton version skew at import time (#11023)
b07e466 Daniel Han committed 2 months ago
Move the Studio MLX pins to mlx 0.32.2 and mlx-vlm 0.7.1 (#11080)
fb4fc4f Long Yixing committed 2 months ago
Studio: do not run Gemma tool call examples written inside code blocks (#11153)
f10b3de Nilay committed 2 months ago
Studio: serve every image an MLX conversation carries, not just the newest (#10970)
004a447 Long Yixing committed 2 months ago
perf(studio): keep MLX VLM generation on one stream (#10962)
b776668 Long Yixing committed 2 months ago
Studio: official image picks default to the hosted FP8 or INT8 checkpoint (#10883)
584b91e Daniel Han committed 2 months ago
Studio: assert the argv, not the call count, in the shutdown spawn-refusal test (#10645)
532485e Daniel Han committed 2 months ago
Release: stop publishing standalone signed scripts (#11123)
82a983c Wasim Yousef Said committed 2 months ago
Stop the empty-logits sentinel from claiming protocol dunders it cannot honour (#11130)
3efb3b8 Daniel Han committed 2 months ago
Keep every nested RoPE base, and never put a per-label mapping in rope_theta (#11192)
14be511 Daniel Han committed 2 months ago
Pin every rank of a distributed launch to its own device, not just quantized ones (#11129)
fa4edb5 Daniel Han committed 2 months ago
Studio: reap the llama-server tree on Windows unload, and unblock a dead spawned backend (#11022)
be395e8 Daniel Han committed 2 months ago
Report the video progress the GPU has actually done (#11004)
6ba103b Daniel Han committed 2 months ago
Installers: require x64 Python on Windows on ARM, and respect active conda (#11024)
37a66c8 Daniel Han committed 2 months ago
Resolve the per layer device through a helper instead of a raw index (#11020)
2a28379 Daniel Han committed 2 months ago
Model picker: collapse a single downloaded quant again (#11190)
e66a7cb Michael Han committed 2 months ago
Studio backend tests: pin the logical CPU count in the oversubscribed-threads spill test (#11120)
653d4ec Daniel Han committed 2 months ago
Resolve apply_qkv and apply_o defensively, and explain zero patch counts (#11038)
ac9f9b8 Daniel Han committed 2 months ago
Studio: use --reasoning where available and name the compressed-tensors refusal (#11072)
e7becaf Daniel Han committed 2 months ago
Keep the RoPE base frequency on transformers 5, and make triton optional (#11037)
65facff Daniel Han committed 2 months ago
Studio: pass search results and documents through to the model on /v1/messages (#11152)
3781136 Nilay committed 2 months ago