Model
The model was trained on carefully curated, high-quality datasets that underwent extensive cleaning and rigorous quality assessment to ensure data consistency, reliability, and relevance. The training pipeline consisted of three stages: pretraining on 500B tokens, SFT on 26M samples, and RL (modified GRPO) across 14 diverse environments. This multi-stage training approach was designed to enhance the model's language understanding, instruction-following capabilities, and ability to perform effectively across a broad range of tasks.
Bielik instruct models have been trained with the use of an original open source framework called ALLaMo implemented by Krzysztof Ociepa.
Model description:
- Developed by: SpeakLeash & ACC Cyfronet AGH
- Language: Multilingual (56 unique languages, optimized for European languages, with a particular focus on Polish)
- Model type: causal decoder-only
- License: Apache 2.0
Bielik-11B-v3.1-Instruct uses its own tool-call and reasoning format
(<tool_call>{"name": …, "arguments": …}</tool_call> and <think>…</think>).
To get structured tool_calls and a separate reasoning field in the
OpenAI-compatible API, start vLLM with the Bielik parsers from
speakleash/bielik-tools.
Requirements: vLLM 0.24 or newer (tested on 0.32). The chat template ships
with the model (chat_template.jinja), so --chat-template is not needed.
git clone https://github.com/speakleash/bielik-tools.git 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 (both register the name
bielik).
--exclude-tools-when-tool-choice-none removes tool definitions from the prompt
when a request sets tool_choice="none"; otherwise the model may still write a
tool call as plain text.
- Older per-version parsers (
*_v0.12.0.py, *_v0.13.0.py, *_v0.15.0.py, the 0.20 plugin)
are frozen and meant for older vLLM releases; see the bielik-tools README.
Reasoning
Thinking is controlled with reasoning_effort: none (default), low, medium,
high, max. With none the model answers directly. The enable_thinking
flag used by Bielik v3.0 is ignored by this template.
from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY") resp = client.chat.completions.create( model="speakleash/Bielik-11B-v3.1-Instruct", messages=[{"role": "user", "content": "Ile jest liczb pierwszych mniejszych od 50?"}], reasoning_effort="high", # or extra_body={"chat_template_kwargs": {"reasoning_effort": "high"}} max_tokens=4096,)msg = resp.choices[0].messageprint("reasoning:", getattr(msg, "reasoning", None) or getattr(msg, "reasoning_content", None))print("answer:", msg.content)
Leave enough max_tokens for reasoning plus the answer: if generation stops while
still thinking, content is empty.
tools = [{ "type": "function", "function": { "name": "get_weather", "description": "Returns the current weather for a city.", "parameters": { "type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"], }, },}] resp = client.chat.completions.create( model="speakleash/Bielik-11B-v3.1-Instruct", messages=[{"role": "user", "content": "Jaka jest pogoda w Krakowie?"}], tools=tools, tool_choice="auto", # "required", "none" or a named function also work)print(resp.choices[0].message.tool_calls)
tool_choice="required" and a named function are enforced by the parser
(constrained decoding of the call; the argument schema itself is not enforced).
- The request-level
tool_choice and response_format do not reach the chat
template. To also add the template's instructions, repeat them in
extra_body={"chat_template_kwargs": {"tool_choice": "required"}}.
- Treat tool arguments as untrusted input.
Default system prompt
The template prepends a built-in system prompt ("Model Guidelines": identity,
knowledge cutoff, current date, safety rules; about 1,000 tokens). Your own system
message is appended under "Additional Instructions". The current date comes from the
server clock; for reproducible runs pass current_date and current_day_of_week
in chat_template_kwargs. To disable the built-in prompt for all requests:
vllm serve … --default-chat-template-kwargs '{"provider_system_prompt": false}'
or per request: extra_body={"chat_template_kwargs": {"provider_system_prompt": False}}.
Limitations and Biases
Bielik-11B-v3.1-Instruct is a quick demonstration that the base model can be easily fine-tuned to achieve compelling and promising performance. It does not have any moderation mechanisms. We're looking forward to engaging with the community in ways to make the model respect guardrails, allowing for deployment in environments requiring moderated outputs.
Bielik-11B-v3.1-Instruct can produce factually incorrect output, and should not be relied on to produce factually accurate data. Bielik-11B-v3.1-Instruct was trained on various public datasets. While great efforts have been taken to clear the training data, it is possible that this model can generate lewd, false, biased or otherwise offensive outputs.
Responsible for training the model
The model could not have been created without the commitment and work of the entire SpeakLeash team, whose contribution is invaluable. Thanks to the hard work of many individuals, it was possible to gather a large amount of content in Polish and establish collaboration between the open-science SpeakLeash project and the HPC center: ACC Cyfronet AGH. Individuals who contributed to the creation of the model:
Sebastian Kondracki,
Marek Magryś,
Sonia Staniek,
Waldemar Boszko,
Maria Czarkowska,
Jacek Chwiła,
Kuba Sołtys,
Dominika Basaj,
Remigiusz Kinas,
Marcin Dąbrowski,
Anna Przybył,
Agnieszka Ratajska,
Daria Bartkowiak,
Witold Wydmański,
Jan Sowa,
Grzegorz Anioł,
Bartosz Bartnicki, and many other wonderful researchers and enthusiasts of the AI world.
Legal Aspects
EU AI Act Transparency Documentation: Bielik 11B v3 EU Public Summary.pdf
Data Protection and Copyright Requests
For removal requests of personally identifiable information (PII) or of copyrighted content, please contact the respective dataset owners or us directly: biuro@speakleash.org.pl.
Citation
Please cite this model using the following format:
@misc{Bielik11Bv31i, title = {Bielik-11B-v3.1-Instruct model card}, author = {Ociepa, Krzysztof and Flis, Łukasz and Gwoździej, Adrian and Wróbel, Krzysztof and Jezierski, Karol and Ciuciura, Igor and {SpeakLeash Team} and {Cyfronet Team}}, year = {2026}, url = {https://huggingface.co/speakleash/Bielik-11B-v3.1-Instruct}, note = {Accessed: 2026-10-31}, % change this date urldate = {2026-10-31} % change this date}
If you have any questions or suggestions, please use the discussion tab. If you want to contact us directly, join our Discord SpeakLeash.
Acknowledgements
We gratefully acknowledge Polish high-performance computing infrastructure PLGrid (HPC Center: ACC Cyfronet AGH) for providing computer facilities and support within computational grant no. [PLG/2026/019559].
This work was supported by the Gaia AI Factory project, funded by the European Union under Grant Agreement No. 101314359 through the EuroHPC Joint Undertaking (EuroHPC JU).