Bielik 7B v0.1
Bielik-7B v0.1 is a 7B-parameter base model built by continuously pretraining Mistral-7B on 70B+ tokens of Polish text, with data quality filtered via an XGBoost classifier. Developed by SpeakLeash with ACK Cyfronet AGH on Polish HPC infrastructure. This is a raw base model — it is not ready for chat or instruction use out of the box.
If you are building a Polish-language fine-tune and want a Mistral-7B-derived starting point with serious Polish pretraining behind it, this is a reasonable choice. Do not deploy it raw — without instruction tuning it will not follow prompts usefully. The 4096 context cap will sting for document-heavy tasks, so check whether that fits your use case before committing VRAM. Hedge: worth it as a fine-tuning base, skip it if you need a ready-to-run chat model.
›Why this rating
Auto-generated rating (Opus 4.7 judge, claude-opus-4-7). Overall 9.15/10. License is explicitly apache-2.0 on the card and correctly flagged commercial-OK. Metadata (7B params, mistral family, Polish base model) matches the card. The description and weaknesses are honest and operator-grade: it flags the base-model trap, the 4096 context limit, and low community traction. Best use case is sharp (Polish fine-tuning base). Minor nit: the description says '70B+ tokens' which matches the card's intro, but the card later clarifies '36 billion tokens for two epochs' — the row's phrasing is defensible but slightly imprecise. Brand fit is modest since this is a base model requiring further training, but the row is upfront about that.
Flags: - Token count phrasing ('70B+ tokens') reflects the card's headline but the card later clarifies 36B tokens × 2 epochs — minor precision concern - Base-only model has narrower brand fit since most runlocalai readers want ready-to-run models, but the row honestly steers users to the Instruct variant
Overview
Bielik-7B v0.1 is a 7B-parameter base model built by continuously pretraining Mistral-7B on 70B+ tokens of Polish text, with data quality filtered via an XGBoost classifier. Developed by SpeakLeash with ACK Cyfronet AGH on Polish HPC infrastructure. This is a raw base model — it is not ready for chat or instruction use out of the box.
Strengths
- Continuously pretrained from Mistral-7B on 70B+ tokens of Polish text
- Training data filtered with an XGBoost quality classifier — not raw web scrape
- Apache-2.0 licence, commercial use permitted
- Solid throughput during training: 9,200+ tokens/GPU/second on Polish HPC
Weaknesses
- Base model only — no instruction tuning or chat fine-tune; needs additional training before deployment
- 4096-token context is short by current standards
- Trained exclusively on Polish; multilingual performance is untested and likely poor
- Low community traction so far (74 likes, 2,789 downloads) — limited third-party evaluation available
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|---|---|
| Q4_K_M | 3.9 GB | 5 GB |
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of Bielik 7B v0.1.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
What's the minimum VRAM to run Bielik 7B v0.1?
Can I use Bielik 7B v0.1 commercially?
What's the context length of Bielik 7B v0.1?
Source: huggingface.co/speakleash/Bielik-7B-v0.1
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify Bielik 7B v0.1 runs on your specific hardware before committing money.