Views
No views yet
speakleash/Bielik-Minitron-7B-v3.0-Instruct do formatu MLX (Apple Silicon), kwantyzacja bf16.| Wariant | Rozmiar | Use case |
|---|---|---|
| MLX-4bit | ~4 GB | Edge / MacBook Air, ograniczona pamięć |
| MLX-6bit | ~5.5 GB | Sweet spot quality/size |
| MLX-8bit | ~8 GB | Wysoka jakość, blisko bf16 |
| MLX-bf16 | ~15 GB | Pełna precyzja, źródło do dalszych konwersji |
pip install mlx-lm1from mlx_lm import load, generate
2
3model, tokenizer = load("agentGreg/Bielik-Minitron-7B-v3.0-Instruct-MLX-bf16")
4
5messages = [
6 {"role": "user", "content": "Wyjaśnij prosto czym różni się prędkość od przyspieszenia."},
7]
8prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
9print(generate(model, tokenizer, prompt=prompt, max_tokens=400, verbose=True))| Pozycja | Model | Wynik |
|---|---|---|
| 🥇 | Bielik-Minitron 7B v3 (MLX 8-bit) | 25/30 (83%) |
| 🥈 | Bielik 4.5B v3 | 24/30 (80%) |
| 🥈 | Gemma 4 E4B (text-only) | 24/30 (80%) |
| 4 | Gemma 4 E4B (multimodal) | 23/30 (77%) |
| 5 | Gemma 3 4B (text-only) | 18/30 (60%) |
| 6 | Gemma 3 4B (multimodal) | 14/30 (47%) |
| 7 | Llama-PLLuM 8B | 3/30 (10%) |
| 7 | PLLuM 12B | 3/30 (10%) |
speakleash/Bielik-Minitron-7B-v3.0-Instruct — © SpeakLeash teammlx-lm (0.31.3)@article{bielik_minitron_2026,
title={Compressing Polish LLMs with Hybrid Pruning and Distillation},
author={SpeakLeash team},
journal={arXiv preprint arXiv:2603.11881},
year={2026}
}