Poland's Bielik 11B v3.1 expands to 56 languages and adds agent tools

SpeakLeash trained version 3.1 on an extra 500 billion tokens. The model now covers every official EU language and, for the first time, languages from outside Europe, including Arabic, Chinese, Japanese and Korean, plus Armenian, Azerbaijani and Georgian[1]. The foundation reports gains of 8 percentage points on the Belebele reading comprehension benchmark and 5 points on INCLUDE, which tests knowledge and reasoning across 44 languages[1].

"We want Bielik to become the model of first choice in Europe, and in time beyond it as well," said Sebastian Kondracki, president of the SpeakLeash Foundation board[1].

The headline change is agentic capability. Bielik 3.1 combines function calling, structured output and an explicit reasoning mode, and it can reason while choosing which tools to call[1]. On tau2-bench, which simulates airline, retail and telecom customer service cases that must be resolved with tools and rules, SpeakLeash says version 3.1 scores 16 to 45 percentage points higher than 3.0 depending on the scenario[1]. These are the developer's own figures and have not been independently reproduced.

The instruct model is on Hugging Face under the Apache 2.0 licence, as a gated repository that asks users to accept access terms[2]. SpeakLeash has also published quantised versions in GGUF, AWQ, FP8, NVFP4 and MLX formats for local and edge deployment[3].

The release is the first of three in three days. Bielik 11B v3.1 DFlash, a lighter draft model that proposes tokens for the main model to verify through speculative decoding, is scheduled for 11 November. Bielik-One, a family of four decision models in different sizes built to classify and score inputs rather than generate text, follows on 12 November[1]. At the time of writing, the DFlash repository was not yet public on SpeakLeash's Hugging Face page[3].

Bielik is one of Europe's most visible sovereign model projects, and permissive licensing plus agent features make it a practical option for teams that want a small multilingual model they can run themselves. The Bielik-One decision models also point to a trend worth watching: small, purpose-built classifiers that sit alongside large generators in production pipelines.