Bielik-One-4B-r1-MLX

MLX conversion based on the model speakleash/Bielik-One-4B-r1.

Run

bfloat16, Apple silicon. systemone_mlx.py (in this repo) serves the bielik-one /v1/systemone API: the official bielik-one frontend, unchanged, over an MLX backend. Decisions only, no chat.

pip install mlx-lm fastapi uvicorn "bielik-one @ git+https://github.com/speakleash/bielik-one@475c896e8ebc964236648c1d1f5bb57268914dae"
hf download agentGreg/Bielik-One-4B-r1-MLX --local-dir Bielik-One-4B-r1-MLX
python Bielik-One-4B-r1-MLX/systemone_mlx.py Bielik-One-4B-r1-MLX        # http://127.0.0.1:8026/v1/systemone, model "decision"

Request and response format, calibration: see the original model card.

Parity with vLLM

Raw probabilities (temperature 1) on the hidden test split of the Tacit benchmark, this conversion on an Apple M5 Max vs the original weights on bielik-one 1.0.8 + vLLM 0.32.0 (bf16, NVIDIA A100). Automated at 5% error: Tacit's calibration fitted on its public dev split, separately for each engine. Public dev split: same answer on 99.65% of 2852 items.

language	items	same answer	largest option-probability difference per item, mean / max	accuracy vLLM / MLX	automated at 5% error, vLLM / MLX
pl	652	99.39%	0.0050 / 0.080	87.6% / 87.9%	67.6% / 68.1%
en	750	99.60%	0.0051 / 0.130	91.1% / 91.2%	65.5% / 66.0%
de	750	99.87%	0.0045 / 0.060	87.1% / 87.1%	68.1% / 68.0%
nb	750	99.33%	0.0055 / 0.062	85.7% / 85.3%	66.3% / 66.7%
all	2902	99.55%		87.9% / 87.9%	66.9% / 67.2%