Orvena 4B, MLX 4-bit
This is the 4-bit MLX build of Orvena 4B, the model Orvena downloads and runs on the iPhone. Orvena 4B is trained for tool use on a phone in 30 languages.
The weights are 4-bit with a group size of 64 (about 4.5 bits per weight including scales), and the vision encoder is included, 3.1 GB in total. The multi-token prediction head is only in the bf16 repository, since Orvena doesn't use it on the phone.
On the Orvena bench of 547 phone tasks in 30 languages, run with this exact build, Orvena 4B passes 440 (80.4%) against 345 (63.1%) for its base model. On an iPhone 17 Pro Max the median task takes 15.1 seconds, against 37.0 seconds for the base model with thinking on auto. The main model card has the full results, and the technical report covers the data, training and evaluation.
Using it
Orvena loads it with mlx-swift-lm's VLMModelFactory, so the folder works as is in Swift apps built on
mlx-swift-lm. The chat template answers without thinking unless
enable_thinking is set to true, which matches how the model was trained. Orvena samples with temperature 0.7 and
top-p 0.8, the values in generation_config.json.
License
Apache-2.0, like Qwen3.5-4B. Built with Qwen. NOTICE lists what was changed.
- Downloads last month
- -
4-bit