Last updated: November 11, 2026
Bielik 11B v3.1, released November 10, 2026 by SpeakLeash and ACK Cyfronet AGH, is an open Apache 2.0 model that expands language support from 32 to 56 and adds function calling, structured output and selectable reasoning effort. A DFlash draft model for faster speculative decoding followed on November 11.
Contents
- What’s new at a glance
- Bielik 11B v3.1 vs v3.0
- How to serve it with tool calling and reasoning
- Speed up generation with DFlash
- Run it locally with GGUF, MLX or FP8
- Limits and what to verify
- Frequently asked questions
Bielik is the open Polish language model family built by the SpeakLeash Foundation with the ACK Cyfronet AGH supercomputing center. Version 3.1 keeps the 11-billion-parameter size of the v3 line but targets agent workloads. The official launch post describes three launches in three days, and the Bielik-11B-v3.1-Instruct model card documents how to use the new features.
What’s new at a glance
| Item | Detail | Source |
|---|---|---|
| Release | Bielik 11B v3.1 on November 10, 2026; DFlash draft model on November 11; Bielik-One decision models scheduled for November 12 | Launch post |
| Size | About 11.17 billion BF16 parameters, causal decoder-only (Llama architecture) | Hugging Face metadata |
| Training | Pretraining on 500B tokens, SFT on 26M samples, RL (modified GRPO) across 14 environments | Model card |
| Languages | 56, up from 32 in v3.0, including Arabic, Chinese, Japanese, Korean, Armenian, Azerbaijani and Georgian | Launch post |
| Agent features | Function calling, structured output, reasoning mode (also used while calling tools) | Launch post |
| Context in the vLLM example | 32,768 tokens (–max-model-len 32768; config max_position_embeddings 32768) | Model card and config |
| License | Apache 2.0 (gated download form on Hugging Face) | Model card |
| Try it | Web chat at chat.bielik.ai | Model card |
Bielik 11B v3.1 vs v3.0
SpeakLeash reports gains over v3.0 on multilingual and agent benchmarks. These are vendor-reported results; we did not re-run them.
| Area | v3.0 | v3.1 |
|---|---|---|
| Supported languages | 32 | 56 |
| Belebele reading comprehension | Baseline | +8 percentage points |
| INCLUDE (44-language knowledge and reasoning) | Baseline | +5 percentage points |
| τ²-bench (airline, retail, telecom agent scenarios) | Baseline | +16 to +45 percentage points, depending on scenario |
| Thinking control | enable_thinking flag | reasoning_effort: none, low, medium, high, max |
The practical change for builders is that v3.1 is meant to complete multi-step tasks with tools rather than only answer questions. For other small open models aimed at agents, see our K-EXAONE 2.5 guide.
How to serve it with tool calling and reasoning
Bielik uses its own formats, <tool_call> for tool calls and <think> for reasoning. To get structured tool_calls and a separate reasoning field through an OpenAI-compatible API, the model card says to use the Bielik parsers:
- Install vLLM 0.24 or newer (SpeakLeash tested 0.32; see our vLLM 0.32 guide). Accept the access form on the Hugging Face model page.
- Clone speakleash/bielik-tools. The chat template ships with the model, so you do not need
--chat-template. - Start the server with
vllm serve speakleash/Bielik-11B-v3.1-Instruct --enable-auto-tool-choice --tool-parser-plugin ./bielik-tools/tools/bielik_vllm_tool_parser.py --tool-call-parser bielik --reasoning-parser-plugin ./bielik-tools/tools/bielik_vllm_reasoning_parser.py --reasoning-parser bielik --exclude-tools-when-tool-choice-none --max-model-len 32768. Load exactly one tool parser and one reasoning parser. - Point an OpenAI client at
http://localhost:8000/v1and setreasoning_efforttolow,medium,highormaxwhen you want thinking. The defaultnoneanswers directly. - Leave enough
max_tokensfor reasoning plus the answer. The card warns that if generation stops while the model is still thinking, the content field is empty. - Pass a
toolsarray withtool_choiceset toauto,required,noneor a named function. Required and named choices are enforced by the parser, but the argument schema itself is not, so validate arguments and treat them as untrusted input.
Two settings are easy to miss. First, request-level tool_choice and response_format do not reach the chat template; repeat them in extra_body={"chat_template_kwargs": {"tool_choice": "required"}} if you want the template’s instructions too. Second, the template adds a built-in system prompt of about 1,000 tokens with identity, knowledge cutoff, current date and safety rules. Your system message is appended under “Additional Instructions.” Turn the built-in prompt off with provider_system_prompt: false, and pass current_date and current_day_of_week for reproducible runs.
Speed up generation with DFlash
Bielik-11B-v3.1-DFlash is a draft model for speculative decoding. It proposes upcoming tokens and the main model verifies them, which SpeakLeash says speeds up output without a loss in quality. It was trained on the same data as the Instruct model and is not usable on its own.
- SGLang (newer than 0.5.6): launch the Instruct model with
--speculative-algorithm DFLASH --speculative-draft-model-path speakleash/Bielik-11B-v3.1-DFlash --speculative-num-draft-tokens 16. Our SGLang 0.5.23 guide covers related serving options. - vLLM (newer than 0.20.0): add
--speculative-config '{"method": "dflash", "model": "speakleash/Bielik-11B-v3.1-DFlash", "num_speculative_tokens": 15}' --attention-backend flash_attn --max-num-batched-tokens 32768.
SpeakLeash also lists DFlash2, FP8-Dynamic and GGUF builds of the draft models. The launch post gives no speed-up figure, so benchmark on your own hardware.
Run it locally with GGUF, MLX or FP8
SpeakLeash published quantized builds alongside the release:
- GGUF for llama.cpp or Ollama: bf16 (about 22.3 GB), Q8_0 (about 11.9 GB), Q6_K (about 9.2 GB), Q5_K_M (about 7.9 GB) and Q4_K_M (about 6.7 GB), per the GGUF repository file list.
- MLX 4-bit and 8-bit builds for Apple silicon.
- FP8-Dynamic, AWQ, NVFP4 and NVFP4A16 builds for GPU serving.
For a quick local test, Q4_K_M is the smallest file. Keep in mind that tool calling with structured output depends on a runtime that parses Bielik’s <tool_call> format; the documented, supported path is vLLM with the bielik-tools parsers. If you want a chat front end over a local server, our Open WebUI 0.13 guide explains the setup.
Limits and what to verify
- The model card states there are no built-in moderation mechanisms and that the model can produce incorrect, biased or offensive output. Add your own filtering for public deployments.
- All benchmark gains above are SpeakLeash-reported comparisons with v3.0.
- The documented serving example uses a 32,768-token context, which is small next to frontier models.
- The DFlash model card lists “32 European languages,” while the Instruct card and launch post say 56. Test the draft model’s acceptance rate on non-European languages.
- Bielik-One, the decision-model family, was scheduled for November 12 and was not yet available when this guide was written.
Frequently asked questions
What is new in Bielik 11B v3.1?
Bielik 11B v3.1, released November 10, 2026 by the SpeakLeash Foundation and ACK Cyfronet AGH, adds 500 billion tokens of training, raises language support from 32 to 56, and adds function calling, structured output and a reasoning mode for agentic tasks.
Is Bielik 11B v3.1 free for commercial use?
Yes. The Bielik-11B-v3.1-Instruct model card lists the Apache 2.0 license, and Bielik’s own guide says Bielik is free of licensing fees for commercial and non-commercial use. The Hugging Face repository asks you to accept a short access form before downloading.
How do I turn on reasoning in Bielik 11B v3.1?
Set reasoning_effort to low, medium, high or max in an OpenAI-compatible request. The default is none, which answers directly. The enable_thinking flag from Bielik v3.0 is ignored by the new chat template.
Does Bielik 11B v3.1 support tool calling in vLLM?
Yes. Start vLLM 0.24 or newer with the bielik tool and reasoning parser plugins from the speakleash/bielik-tools repository and –enable-auto-tool-choice. You then get structured tool_calls and a separate reasoning field in responses.
What is Bielik DFlash?
Bielik-11B-v3.1-DFlash is a draft model for speculative decoding, released November 11, 2026. It proposes upcoming tokens that Bielik-11B-v3.1-Instruct verifies, which speeds up generation. It is not a standalone model and must be paired with the Instruct model.
Sources
- Launching Bielik 11B v3.1, DFlash and the Bielik-One family (Bielik.ai, November 10, 2026)
- speakleash/Bielik-11B-v3.1-Instruct model card (Hugging Face)
- speakleash/Bielik-11B-v3.1-DFlash model card (Hugging Face)
- speakleash/Bielik-11B-v3.1-Instruct-GGUF (Hugging Face)
- speakleash/bielik-tools (GitHub)
Image: Official Bielik 11B v3.1 launch graphic from Bielik.ai, used for editorial coverage of the November 2026 release. Bielik is a project of the SpeakLeash Foundation; FreeAnswers is not affiliated with SpeakLeash.