Bielik 11B v3.1: Polish AI Model Now Supports 56 Languages
The Polish open model Bielik 11B v3.1 now handles 56 languages, up from 32, and adds agentic skills such as function calling. A 100B+ model is due on 26 November.
SUMO· the number puzzle
Dialo· learn languages by speaking
ediviewer· read EDI files like documents
giftdetect· the right gift in minutesThe Polish language model Bielik has a new version. Bielik 11B v3.1 supports 56 languages, up from 32, handles complex tasks better and can use external tools. The SpeakLeash Foundation and the Academic Computer Centre Cyfronet AGH are behind the release.
The model was trained on an additional 500 billion tokens, and the team says it kept its strong Polish skills. It covers all official languages of the European Union, plus Arabic, Chinese, Japanese and Korean, as well as Caucasian languages such as Armenian, Azerbaijani and Georgian. On the Belebele reading-comprehension benchmark the score rose by 8 percentage points, and on INCLUDE, which tests knowledge and reasoning in 44 languages, by 5 points.
Agentic skills and new models
A key change is the focus on agentic abilities: the model can plan steps and use external tools to complete more complex tasks. It supports function calling, structured output and a reasoning mode. In τ²-bench, which simulates customer-service cases in telecoms, retail and airlines, Bielik 3.1 beat its predecessor by 16 to 45 percentage points depending on the scenario.
The launch also introduces the Bielik-One family: four decision models that, instead of generating text, assign scores or probabilities to information. They suit email sorting, ticket classification, content moderation and document checks. Automation thresholds can be set so that clear cases are handled automatically while ambiguous ones go to a human. The models come in different sizes to match available computing power.
A third release, Bielik 11B v3.1 DFlash, is an auxiliary model that speeds up generation through speculative decoding: it proposes text fragments that the main model then verifies, which the team says does not lower answer quality. All models and weights are to be published under open science principles within SpeakLeash.
A 100-billion-parameter model next
At the Bielik Summit 2026 on 26 November, the creators plan to present details of the first Bielik with more than 100 billion parameters. It is to be multimodal and built from scratch on a proprietary architecture. Work has been under way for six months, with support from Polish companies and institutions. The team notes that open multimodal models of that scale are practically absent in Europe.
Apps worth a look
SUMO: the number puzzleTap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.
Dialo: learn languages by speakingTalk to an AI conversation partner; every sentence comes back corrected and explained in your language. 9 languages, A1 to C2.
ediviewer: read EDI files like documentsOpen EDIFACT and X12 messages, see every segment explained, convert to JSON/XML, export PDF and Excel. Runs locally, no upload.
giftdetect: the right gift in minutesPick who it is for, the occasion and their personality. The search engine reads product descriptions from local shops and ranks what fits. Free, no sign-up.
