SpeakLeash Releases Bielik 11B v3.1 in 56 Languages

SpeakLeash and ACK Cyfronet AGH have released Bielik 11B v3.1, an 11-billion-parameter open model from Poland that now covers 56 languages and adds function calling, structured output and a reasoning mode. In a blog post dated 10 November 2026, the foundation says it “trained Bielik 11B v3.1 on an additional 500 billion tokens. This raised the number of supported languages from 32 in version 3.0 to 56.” The weights are published on Hugging Face under the Apache 2.0 licence.

The same post sets out a three-day schedule: the 11B v3.1 instruct model on 10 November, a DFlash speculative-decoding draft model on 11 November, and the Bielik-One family of four decision models on 12 November. Bielik-One has been announced and is not released. At the time of writing, on the morning of 11 November, no DFlash repository had appeared under SpeakLeash on Hugging Face either, so that model should also be treated as planned rather than already out.

Bielik 11B v3.1 now covers every official EU language, plus four from further afield

The blog says users can now talk to Bielik “in every official language of the European Union.” For the first time it also supports languages from outside Europe, including Arabic, Chinese, Japanese and Korean, plus Armenian, Azerbaijani and Georgian. The model card describes a three-stage pipeline: pretraining on 500 billion tokens, supervised fine-tuning on 26 million samples, and reinforcement learning across 14 environments. Training ran on the Athena and Helios supercomputers and in the Gaia AI Factory, “funded by the European Union through the EuroHPC Joint Undertaking.”

Sebastian Kondracki, president of the SpeakLeash Foundation, said: “We want Bielik to become the model of first choice in Europe, and in time beyond it as well.” That is an ambition, not a market result. The card is more cautious in another respect: “It does not have any moderation mechanisms.” For a model that can now call tools and return structured output, that warning is part of the release, not a footnote to skip.

The agentic additions are the other half of the story. The blog says version 3.1 combines function calling, structured output and a reasoning mode that, for harder tasks, switches to an explicit thinking pass. SpeakLeash reports its own benchmark gains against version 3.0: eight percentage points on Belebele, five on INCLUDE, and “16 to 45 percentage points higher” on τ²-bench, depending on the scenario. Those figures are the lab’s comparisons with its previous model, not an independent ranking against the wider field.

A European open model is not the same thing as a published decision-model family

Poland’s release sits in a familiar European pattern. Domyn opened Edge, a 4-billion-parameter agentic model trained on Italy’s Leonardo supercomputer. Switzerland’s Giotto lab released Alto and Solo as sovereign AI. Mistral has opened the weights of Large 4. Each of those stories is about a lab putting a model in public. None of them makes Bielik-One available a day early, and none of them supplies the moderation the Bielik card says is missing.

The EU funding line is real context: Gaia and EuroHPC paid for compute. It is not a safety certification, and it is not a promise that the 12 November decision models will ship on time or in the form described. The blog says Bielik-One’s four models “are not built to generate text” but to classify and score against criteria set in advance, for routing, screening and similar tasks. That is a planned product family. It should be judged when the weights appear.

Until then, the model that is actually out is Bielik 11B v3.1 Instruct: 56 languages, Apache 2.0, tool use, and a card that tells users it has no moderation mechanisms. The next date on SpeakLeash’s own calendar is 12 November, when Bielik-One is due. The question worth asking on that day is whether the decision models are on Hugging Face, not whether the 10 November instruct model has been retroactively relabelled.

Sources

1. SpeakLeash, “Launching Bielik 11B v3.1, DFlash and the Bielik-One family,” 10 November 2026. https://www.bielik.ai/en/blog/launching-bielik-11b-v31-dflash-and-the-bielik-one-family
2. Hugging Face, speakleash/Bielik-11B-v3.1-Instruct model card. https://huggingface.co/speakleash/Bielik-11B-v3.1-Instruct

Related stories

Domyn’s open Edge model, trained on Italy’s Leonardo supercomputer, is the nearest recent TSN account of another European lab releasing an open agentic model from a public HPC centre.

Also relevant:

• Swiss Lab Giotto Releases Alto and Solo Models as Sovereign AI
• Mistral Opens Large 4 Weights: a 1-Trillion-Parameter Model