Attribute	Value
Base model	speakleash/Bielik-11B-v2.3-Instruct
Architecture	Mistral (50 layers, 4096 hidden, 32 heads, 8 KV heads)
Quantization method	QuIP# with E8P12 lattice codebook
Precision	2-bit weights (FP16 base)
Model size	3.26 GB (vs ~22 GB FP16, ~6.7x compression)
Calibration	CulturaX-PL (512 samples, 4096 tokens each)

Evaluation (Polish LLM Leaderboard)

Evaluated on 22/23 tasks from the SpeakLeash Open PL LLM Leaderboard (eq_bench excluded due to private dataset).

Metric	Score
Normalized avg (22 tasks)	61.10
FP16 baseline	65.71
Retention	~93% of FP16 quality

Full per-task results: Jakubrd4/bielik-q2-sharp

Comparison with IQ2_XXS (llama.cpp)

Metric	QuIP# E8P12	IQ2_XXS	FP16
Raw avg (22 tasks)	71.92	72.07	75.40
Tasks won (head-to-head)	11/22	11/22	—

QuIP# achieves parity with llama.cpp IQ2_XXS on 22 Polish benchmarks (delta -0.15%).

Usage

Requires quip-sharp for inference:

python

from lib.utils.unsafe_import import model_from_hf_path

model, tokenizer = model_from_hf_path(
    "Jakubrd4/Bielik-11B-v2.3-Instruct-QuIP-2bit"
)

Note: Bielik uses Mistral architecture. QuIP# expects LlamaConfig, so a patch in model_from_hf_path() is needed to convert MistralConfig to LlamaConfig (map sliding_window -> None, attention_dropout -> 0).

Related Resources

Evaluation data & scripts: Jakubrd4/bielik-q2-sharp
Hessian matrices: Jakubrd4/bielik-quip-e8p12 (in hessians/ directory)
Base model: speakleash/Bielik-11B-v2.3-Instruct
QuIP# paper: arXiv:2307.13304