Mirai's Qwen3.5-2B Small Quantization
A small, high-quality quantization of Qwen3.5-2B prepared for efficient local inference on Apple silicon.
Evaluated with romakl on its default set: 20 conversations and 15,438 teacher-generated continuation tokens. This checkpoint: KL 0.2406042; top-1 agreement 78.33%. The plot shows measured checkpoints within ±50% in token-generating weight size, without clipping the selected KL values. The three Bonsai 2 packings were evaluated separately with the Prism runtime. Vision, MTP and speculator weights are excluded from size. KL measures deviation from the base teacher.
Quickstart
If you are on macOS, the easiest way is to install the mirai Homebrew package and then run the CLI:
brew install mirai
mirai --model trymirai/Qwen3.5-2B-S
Currently only Apple silicon inference is supported. If you want to build things from source, read this overview.
Method
Mirai Small uses QTIP-style trellis-coded quantization at about 2.5 bits per weight, with the bits of each matrix chosen from measured KL ablations. Kronecker-structured random rotations are used to reduce weight outliers, and rounding is weighted by a Fisher sketch of the model's own output distribution. The checkpoint was prepared with post-training quantization followed by quantization-aware distillation of the scales.
Citation
If you find our work helpful, feel free to give us a cite.
@misc{mirai-quant,
title = {{Mirai Quantization}: Redefining the speed-quality frontier for local LLMs on Apple silicon},
author = {Artur Chakhvadze and Ryan Mathieu and Roman Knyazhitskiy and Nikolai Voinilenko and Chen-Chen Yeh and Artur Mullakhmetov and Eugene Bokhan and others},
note = {In collaboration with others at Mirai Labs},
month = {June},
year = {2026},
url = {https://trymirai.com/blog/quantization}
}
Original model
This is a quantized version of Qwen/Qwen3.5-2B. For architecture details, intended use, evaluations, and limitations, see the original model card.