News
Bielik 11B v3.1 Adds 56 Languages and Agent Skills, Team Teases 100B+ Parameter Model
The SpeakLeash Foundation and Cyfronet AGH have released Bielik 11B v3.1. After further training on 500 billion tokens, the model supports 56 languages instead of 32 and handles agentic tasks better. The team also announced the first Bielik with more than 100 billion parameters.
The Polish language model Bielik has a new version. Bielik 11B v3.1, unveiled on November 10 by the SpeakLeash Foundation and the Academic Computer Centre Cyfronet AGH, supports 56 languages instead of the previous 32 and gains features useful for building AI agents.
What has changed
According to reports by ITwiz, Telepolis and MyCompanyPolska, the model underwent additional training on 500 billion tokens. The language list covers all official languages of the European Union, as well as Arabic, Chinese, Japanese and Korean, and Armenian, Azerbaijani and Georgian.
The developers say quality in Polish and knowledge of local context did not suffer from the expansion. That conclusion comes from their own materials, and the sources contain no independent verification.
The team reports benchmark results: Belebele, which measures reading comprehension, rose by 8 points, and INCLUDE, which tests knowledge and reasoning in 44 languages, rose by 5 points compared with version 3.0.
Agentic skills
The new version combines function calling, structured output and explicit reasoning. According to ITwiz, reasoning also works while tools are being called, which is meant to help with multi-step tasks.
On τ²-bench, which simulates customer service in airlines, retail and telecommunications, Bielik 11B v3.1 beat its predecessor by 16-45 percentage points, depending on the scenario. The result is reported by the developers.
In many business processes we don't need an elaborate answer from the model, but a specific result, for example assigning a ticket to the right category or scoring it on a defined scale. - Krzysztof Wróbel, Principal AI Researcher at the SpeakLeash Foundation (ITwiz, MyCompanyPolska)
Bielik-One and DFlash
Bielik-One is a family of four decision models of different sizes. Instead of generating long text, they classify and score information, for example when sorting email, handling customer tickets, moderating content or checking document compliance.
According to ITwiz, the models return probability distributions, so thresholds can be set: clear-cut cases are automated, doubtful ones go to a human.
Bielik 11B v3.1 DFlash is a helper model for speculative decoding. A lighter model proposes chunks of text and the main model verifies them, which speeds up generation without lowering quality, the developers say.
Plans for a 100B+ model
For six months the team has been working on the first Bielik with more than 100 billion parameters. It is meant to be multimodal and built from scratch on a proprietary architecture, with support from Polish companies and institutions. Details are due at Bielik Summit 2026 on November 26; MyCompanyPolska notes that no release date or cost has been given.
We want to change that. - Sebastian Kondracki, president of SpeakLeash, on the fact that in Europe only Mistral effectively builds open models of this class (MyCompanyPolska)
The models and their weights are to be released on open science terms under the SpeakLeash project. The sources do not name a license, so commercial-use terms should be checked when the weights are published. For Polish companies, according to MyCompanyPolska, the approach matters because it lets them run the models on their own infrastructure.
