Picked for your PC
Ornith 1.5 9B
Runs well
A coding agent model on a Qwen 3.5 base. It reasons, then edits. Text only. 256K context.
~5.9 GB~54 tok/s
Which model can your PC run?
Osaurus for Windows runs open models on your graphics card, with CUDA on NVIDIA and Vulkan on AMD and Intel. Pick your card and memory to see which models fit and roughly how fast they run. No graphics card? Cloud models work on any Windows 11 PC.
Your PC
5 run well7 tight fit1 too largeof 13 curated models · ~10 GB usable for models
Gemma 4 12B
Runs well
Google's 12B model for everyday chat, with tool calling, a thinking mode, and image understanding. 256K context.
~8.0 GB~44 tok/sVision
LFM2.5 8B-A1B
Runs well
A Liquid AI mixture-of-experts model with about 1B parameters active per token, for the fastest local chat. Its license limits commercial use above $10M in yearly revenue.
~6.1 GB~177 tok/s
Raptor 0.6 4B
Runs well
Osaurus's own 4B agent model, tuned for Osaurus tools. It reasons by default, and it starts chatting in seconds while a bigger model downloads.
~4.0 GB~86 tok/s
Qwen3-VL 8B
Runs well
A vision-language model that reads images and chats about them. 256K context.
~7.7 GB~62 tok/sVision
gpt-oss 20B
Tight fit
OpenAI's open reasoning model, a mixture of experts with about 3.6B parameters active per token. Reasoning effort is a setting. 131K context.
~11.9 GB~78 tok/s
Qwen 3.6 35B-A3B
Tight fit
A 35B mixture-of-experts model with about 3B parameters active per token, so it answers like a large model at the speed of a small one. Reads images. 256K context.
~21.8 GB~42 tok/sVision
Ornith 1.5 35B-A3B
Tight fit
The larger Ornith coding agent, a mixture of experts with about 3B parameters active per token. Text only. 256K context.
~20.8 GB~45 tok/s
Gemma 4 31B
Tight fit
The largest dense Gemma 4, with image understanding. 256K context.
~20.9 GB~5.7 tok/sVision
Qwen3 Coder 30B-A3B
Tight fit
A mixture-of-experts model built for code, with about 3B parameters active per token. 256K context.
~18.2 GB~53 tok/s
Qwen 3.8 27B
Tight fit
A dense 27B agent model that also reads images. The strongest all-round model in this list that fits a 24 GB card. 256K context.
~17.2 GB~7.4 tok/sVision
Gemma 4 26B-A4B QAT
Tight fit
A Gemma 4 mixture of experts trained to hold up at 4 bits, with about 4B parameters active per token. Reads images. 256K context.
~15.3 GB~56 tok/sVision
gpt-oss 120B
Too large
OpenAI's large open reasoning model, with about 5.1B parameters active per token. For workstations with a lot of memory. 131K context.
~60.0 GB
Memory is the app's own estimate at the context it starts a model with: 16K below 12 GB, 32K on 12 to 19 GB cards, and 64K on larger cards and unified memory. Speed is a rough estimate from the card's memory bandwidth. Real numbers vary with the driver, laptop or desktop parts, and what else your PC is doing.