Picked for your Mac
Raptor 0.6
Runs well
Osaurus's own 4B agent model. 3.7 GB, built on Spark-X2.5-4B, and tuned for Osaurus tools. Built for 16 GB Macs, and runs on as little as 8 GB.
~3.4 GB~21 tok/s
Which model can your Mac run?
Osaurus runs on Apple Silicon with macOS 15.5 or later. Pick your chip and memory to see which models fit and roughly how fast they run.
Your Mac
11 run well4 tight fit42 too largeof 57 curated models · ~11 GB usable for models
Gemma 4 E2B
Runs well
The smallest high-precision multimodal model in this list. Images, video, and audio. 128K context.
~6.9 GB~39 tok/s
Bonsai 27B 1-bit
Runs well
The smallest Bonsai build, at about 4.7 GB. A 27B vision model in 1-bit weights.
~5.4 GB~23 tok/sVision
Gemma 4 E4B
Runs well
The smallest E4B download. Lower precision, for Macs with less memory.
~8.0 GB~39 tok/sVision
Raptor 0.5
Runs well
Osaurus's earlier 8B-A1B agent model. 6.3 GB, tuned for Osaurus tools, and built to stay loaded on 8–16 GB Macs. Replaced by Raptor 0.6. 131K context.
~6.3 GB~78 tok/s
Gemma 4 E4B QAT
Runs well
A small multimodal model in a 4-bit build trained to hold up at that size. 128K context.
~6.9 GB~39 tok/sVision
Z-Image Turbo
Runs well
A quick text-to-image model from Tongyi-MAI that makes an image in a few steps. Runs locally.
~6.9 GBText to image
ZAYA1 8B
Runs well
A compact model for reasoning and tool use.
~6.8 GB~20 tok/s
ZAYA1 8B
Runs well
The same reasoning and tool-use model with 4-bit TurboQuant experts.
~5.8 GB~20 tok/s
Gemma 4 E2B
Runs well
The smallest Gemma 4 download, for Macs with the least memory.
~5.1 GB~78 tok/s
Gemma 4 E2B QAT
Runs well
A 4-bit build trained to hold up at that size. The smallest multimodal model in this list. 128K context.
~4.7 GB~78 tok/s
Gemma 4 E4B
Tight fit
The E4B build to try first. Images, video, and audio. 128K context.
~10.5 GB~20 tok/sVision
Bonsai 27B Ternary
Tight fit
A 27B vision model compressed to about 8 GB with ternary weights.
~9.4 GB~11 tok/sVision
LFM2.5 8B
Tight fit
A Liquid AI mixture-of-experts model with about 1B parameters active per token. Quick chat at high precision. 128K context.
~10.2 GB~78 tok/s
Gemma 4 12B QAT
Tight fit
The 12B multimodal model in a 4-bit build trained to hold up at that size. 128K context.
~9.2 GB~13 tok/sVision
Ornith 1.0 35B
Too large
A vision-language mixture-of-experts model for coding agents, at high precision. 256K context.
~42.7 GBVision
Gemma 4 12B
Too large
Google's multimodal model for images, video, and audio, at high precision. 128K context.
~31.2 GBVision
Ornith 1.0 9B
Too large
A vision-language model tuned for coding agents, at high precision. 256K context.
~11.8 GBVision
Kimi K2.6
Too large
A very large vision model in a K-quant TurboQuant build.
~410.2 GBVision
MiniMax M2.7
Too large
A 228B mixture-of-experts agent model with 2-bit TurboQuant experts, for the least memory while it runs. 192K context.
~141.3 GB
MiniMax M2.7
Too large
A 228B mixture-of-experts agent model with 4-bit TurboQuant experts. Close to full-precision quality in about a quarter of the disk space. 192K context.
~136.1 GB
Hunyuan 3 Preview
Too large
A preview of a 295B mixture-of-experts model, in a K-quant TurboQuant build.
~126.9 GB
Nemotron-3 Ultra 550B
Too large
NVIDIA's 550B mixture-of-experts reasoning model, with about 55B parameters active per token. For Macs with the most memory.
~123.0 GB
Allosaurus v0.1
Too large
Qwen3.8-Flash-Next 125B, calibrated for Osaurus tool use, coding, and reasoning. About 35 GiB in memory, so it needs a 64 GB Mac to run at full speed. Its 54 GiB n-gram table stays on the SSD. Text, images, and video.
~35.0 GB
Allosaurus v0.2
Too large
Qwen3.8-Flash-Next 125B, compressed further to fit a 48 GB Mac. About 27 GiB in memory. Its 54 GiB n-gram table stays on the SSD. Text only.
~27.0 GB
DeepSeek V4 Flash
Too large
A large reasoning model in a K-quant TurboQuant build.
~100.0 GB
Mistral Medium 3.5 128B
Too large
A 128B model with Pixtral vision. The faster build. 256K context and 24 languages.
~99.8 GBVision
Step 3.7 Flash
Too large
A vision-language model in a JANG K-quant build.
~92.8 GBVision
DeepSeek V4 Flash
Too large
A large reasoning model in a 2-bit TurboQuant build.
~92.7 GB
Ling 2.6 Flash
Too large
A BailingHybrid mixture-of-experts model. The higher-quality Ling build.
~78.3 GB
Ideogram 4
Too large
Ideogram 4 in a community build, known for clean lettering.
~64.1 GBText to image
Mistral Medium 3.5 128B
Too large
The same 128B model with a 2-bit TurboQuant text decoder, at about 41 GB. 256K context.
~47.5 GBVision
MiniMax M2.7 Small
Too large
A smaller mixture-of-experts agent model with TurboQuant experts. 192K context.
~44.7 GB
Qwen 3.6 35B MoE
Too large
A 35B mixture-of-experts vision model with about 3B parameters active per token. Higher precision than the MXFP4 build, with faster replies from predicting several tokens at a time. 256K context.
~43.7 GBVision
Qwen-Image-Edit
Too large
Qwen-Image tuned for editing. Change, restyle, or add to an image from a text instruction. Runs locally.
~43.6 GBImage edit
Qwen-Image
Too large
Qwen's 20B text-to-image model, good at writing text inside images. Runs locally.
~42.0 GBText to image
Qwen-Image-Edit
Too large
Qwen-Image tuned for editing. Change, restyle, or add to an image from a text instruction. Runs locally.
~37.7 GBImage edit
Ideogram 4
Too large
Ideogram 4 in a community build, known for clean lettering.
~37.5 GBText to image
Qwen-Image
Too large
Qwen's 20B text-to-image model, good at writing text inside images. Runs locally.
~36.1 GBText to image
Ling 2.6 Flash
Too large
The same Ling model with TurboQuant experts, for less memory.
~35.6 GB
Qwen-Image-Edit
Too large
Qwen-Image tuned for editing. Change, restyle, or add to an image from a text instruction. Runs locally.
~34.7 GBImage edit
Qwen 3.6 27B
Too large
A high-precision build that predicts several tokens at a time to speed up replies. 256K context.
~33.9 GBVision
Qwen-Image-Edit
Too large
Qwen-Image tuned for editing. Change, restyle, or add to an image from a text instruction. Runs locally.
~31.7 GBImage edit
Qwen-Image
Too large
Qwen's 20B text-to-image model, good at writing text inside images. Runs locally.
~30.2 GBText to image
Nemotron-3 Nano 30B
Too large
An NVIDIA reasoning model that mixes Mamba-2 and mixture-of-experts layers. The fastest build of the three. 262K context.
~26.4 GB
Laguna XS.2
Too large
Poolside's 33B mixture-of-experts model for coding agents, with 3B parameters active per token. The faster build. 131K context.
~24.4 GB
Nemotron-3 Nano 30B
Too large
The same reasoning model with 4-bit TurboQuant experts, close to full-precision quality. 262K context.
~23.2 GB
Holo3 35B MoE
Too large
The same on-screen agent model with 4-bit TurboQuant experts.
~22.9 GB
Gemma 4 31B QAT
Too large
The largest dense Gemma 4, in a 4-bit build trained to hold up at that size. 128K context.
~22.8 GBVision
Holo3 35B MoE
Too large
A vision mixture-of-experts model that operates apps on screen.
~22.5 GB
Qwen 3.6 35B MoE
Too large
A 35B mixture-of-experts vision model in a smaller MXFP4 build. 256K context.
~22.5 GBVision
Gemma 4 12B
Too large
A smaller, lower-precision build of the 12B model. 128K context.
~18.5 GBVision
Gemma 4 26B-A4B QAT
Too large
A vision mixture-of-experts model with about 4B parameters active per token, in a 4-bit build trained to hold up at that size. 128K context.
~18.2 GBVision
Qwen 3.6 27B
Too large
A dense vision model, and the most-downloaded model in this list. 256K context.
~17.7 GBVision
Nemotron-3 Nano 30B
Too large
The same reasoning model with 2-bit TurboQuant experts, for the least memory. 262K context.
~14.7 GB
Z-Image Turbo
Too large
A quick text-to-image model from Tongyi-MAI that makes an image in a few steps. Runs locally.
~12.8 GBText to image
Laguna XS.2
Too large
The same coding model with 2-bit TurboQuant experts, at about 10 GB. 131K context.
~11.8 GB
Speed is a rough estimate from the chip's memory bandwidth. Real numbers vary with context length, quantization, and what else your Mac is doing.