Models
Models
Open models from Osaurus that run on your own computer, built for steady tool use. Each one says whether it runs on Mac, Windows, or both. Osaurus also runs other open models from Hugging Face on either one.
Not sure what your computer can run? See which models fit your Mac or PC.
Allosaurus v0.2
Qwen's 125B Qwen3.8-Flash-Next, compressed to fit a 48 GB Mac. About 27 GiB in memory, with its 54 GiB lookup table on the SSD. 93% correct on everyday tasks, and tool-calling turns up to 36% faster with speculative decoding.
125B
Parameters, 6.7B active
27 GiB
In memory
48 GB
Minimum Mac memory
Raptor 0.6
A 4B dense agent model built on Spark-X2.5-4B, tuned for Osaurus tools, and compressed to 3.7 GB. It reasons by default, writes about 105 tokens per second on an M5 Max, and runs in as little as 8 GB of memory. On Windows it ships as a GGUF.
4.1B
Dense parameters
3.7 GB
Bundle size
105 tokens/s
Speed, M5 Max
Raptor 0.5
An 8B mixture-of-experts model that uses about 1B parameters per token. It runs every Osaurus tool in 6.3 GB and stays loaded on the 8–16 GB Macs most people own. Replaced by Raptor 0.6.
7.9B
Total parameters
~1B
Active per token
6.3 GB
Bundle size
AppleScript-8B & AppleScript-16B-A4B
Two open models that turn plain-English requests into working macOS automation. They run on your Mac, call tools, and are compressed to fit Apple Silicon.
8B & 16B
Model sizes
8–16 GB
RAM required
Local
Private by default