SSyncDev

Bielik-11B-v3.1-Instruct

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Bielik-11B-v3.1-Instruct is SpeakLeash's open Polish-first 11B instruct model, trained on ~56 languages, with five reasoning levels and tool calling.

Key specifications

Specification score
30.0
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input

Produces

  • Text output

Capabilities

ReasoningFunction callingOpen weights

What is Bielik-11B-v3.1-Instruct?

Polish is a hard language for models trained mostly on English: seven grammatical cases, rich inflection, word order that moves. Bielik-11B-v3.1-Instruct comes from the SpeakLeash project and ACC Cyfronet AGH, and it is trained with Polish and other European languages at the front of the queue.

It is an 11B decoder-only model on the llama architecture, built on Bielik-11B-v3-Base, released under Apache-2.0 on 10 November 2026.

New in 3.1: a reasoning dial

Version 3.1 adds optional reasoning through a reasoning_effort setting with five levels: none (the default), low, medium, high and max. The old enable_thinking flag from v3.0 is ignored. Give reasoning enough max_tokens, or content can come back empty because the budget ran out inside the thinking.

Tool calls need its own parsers

Structured tool calls arrive in <tool_call> blocks and reasoning in <think> blocks. To get them as proper tool_calls and a separate reasoning field, vLLM needs SpeakLeash's custom bielik parsers from its tools repository.

  • vLLM 0.24 or newer (tested on 0.32); the chat template ships with the model.
  • Load exactly one tool parser and one reasoning parser, both named bielik.
  • Use --exclude-tools-when-tool-choice-none so the model does not write tool calls as plain text when tools are off.
  • required and named-function choices are enforced, argument schemas are not. Validate arguments yourself.

A hidden system prompt

This catches people out. The template prepends a built-in guidelines prompt of about 1,000 tokens covering identity, knowledge cutoff, date and safety, and your system message is appended under it. Dates come from the server clock unless you pass them. You can switch it off with provider_system_prompt: false. Budget for those tokens, or disable them if you supply your own.

Facts and gaps

Item Detail
Parameters ~11B, BF16
Training 500B tokens pretraining, 26M SFT samples, GRPO-style RL across 14 environments
Languages 56 in the description; 58 in the metadata tags
Context Not stated; the vLLM example uses 32,768
Benchmarks None published for 3.1
Access Public, but you must share your Hugging Face details to download

With no benchmarks published, the language-count mismatch and the unstated context length are both things to settle in your own tests.

Is it the right Polish model?

The model has no built-in moderation, so add your own. Then try it on what you actually ship: Polish customer emails, legal summaries, public-sector forms. Run 50 prompts at effort none and at medium and compare quality and latency against Mistral Small 4. For translation-heavy work, include Command A Translate in the comparison. More options are under translation.

How to evaluate Bielik-11B-v3.1-Instruct

Compare the model on the work you need to complete.

Tasks and inputs

Bielik-11B-v3.1-Instruct is listed for Text generation, Translation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

none (default), low, medium, high and max, set through reasoning_effort.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Tencent

Hy4 Flash

Tencent Hy4 Flash is the cost-focused companion to Hy4, keeping vision and agent skills at a far lower token price. Availability is still unsettled.

1.05M context

Alibaba

Qwen4 Max

Qwen4 Max (qwen4-max) is Alibaba's first Qwen4-generation flagship: 1M-token context, text, image and video input, thinking on by default, $1.50/$4.50 per 1M.

1M context

Sarvam

Sarvam-105B-1711

Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.

— context

IBM

Granite 4.1 8B LLMON Preview

Granite 4.1 8B LLMON Preview is IBM's model trained on block-structured LLMON prompts to resist indirect prompt injection in agents and RAG.

— context

NVIDIA

Nemotron-Labs Sigma 8B

Nemotron-Labs Sigma 8B is NVIDIA's open continuous diffusion language model, denoising 64-token blocks in parallel; English, GPU-only, 8K context.

8.19K context

Alibaba

Qwen4 Max Preview

Qwen4 Max Preview was Alibaba's five-day early-access build of its Qwen4 flagship, offered only through a Token Plan subscription before GA qwen4-max.

1M context