Bielik 11B v3.1,
on a GPU of your own.
The newest Bielik — the open Polish language model from SpeakLeash and ACK Cyfronet AGH, released on 10 November 2026. Served with vLLM on a dedicated GPU in Poland, behind an OpenAI-compatible /v1 endpoint.
- 11.2B
- 32,768 tokens
- Apache 2.0
- BF16 · FP8 · NVFP4
- vLLM · /v1
The model from the published weights
- Developer
- SpeakLeash · ACK Cyfronet AGH
- Released
- November 10, 2026
- Parameters
- 11,168,796,672 (11.2B)
- Licence
- Apache 2.0
- Context
- 32,768 tokens
- Languages
- 57 language tags · optimised for Polish
- Architecture
- Llama-type decoder · 50 layers
- Attention
- GQA · 32 query / 8 KV heads · head dim 128
- Vocabulary
- 32,128 tokens
- Chat template
- tool calls · reasoning (think) mode
- Draft model
- DFlash · 1,174,455,552 (1.2B) parameters · speculative decoding
Which GPU fits tokens of KV cache left
| Weights | Size | RTX 5090 · 32 GB | DGX Spark · 128 GB | H200 · 141 GB |
|---|---|---|---|---|
| BF16 | 22.34 GB | 31,542under one full-context sequence | 453,41713 × full-context sequences | 510,54615 × full-context sequences |
| FP8 dynamic | 11.44 GB | 84,7652 × full-context sequences | 506,64015 × full-context sequences | 563,76917 × full-context sequences |
| NVFP4 | 6.66 GB | 108,1053 × full-context sequences | 529,98016 × full-context sequences | 587,10917 × full-context sequences |
| AWQ 4-bit | 6.19 GB | 110,4003 × full-context sequences | 532,27516 × full-context sequences | 589,40417 × full-context sequences |
Memory vLLM takes (90% of the GPU) minus the weights, divided by 204,800 B of BF16 KV cache per token (2 × 50 layers × 8 KV heads × 128 × 2 B). One full 32,768-token conversation needs one sequence.
How it is served one tenant per GPU
- Endpoint
- OpenAI-compatible /v1 — chat, completions, streaming.
- Engine
- vLLM, the weight format of your choice: BF16, FP8 or NVFP4.
- Speed-up
- Optional DFlash speculative decoding with the official Bielik draft model.
- Tenancy
- One customer per GPU — nobody else’s requests run on it.
- Location
- Poland, EU — on machines we own and run.
- Billing
- Per hour of GPU time, USD + VAT. Longer blocks quoted on request.
Run Bielik on your own GPU.
Write what load you expect and how much context you need — the weight format and the machine follow from that. Reply within one business day.