This week belongs to Bielik. Together with ACK Cyfronet AGH, we are releasing Bielik 11B v3.1: it speaks 56 languages (including languages from outside Europe), can call external tools and handles tasks made up of many steps. It comes with two new companion models: Bielik-One, the first decision models in the Bielik family, and Bielik DFlash, an accelerator. Three launches in three days. Here is everything you need to know.
Launch schedule
| Date | Launch |
|---|---|
| 10 November 2026 (Tuesday) | Bielik 11B v3.1: a new version of our flagship model |
| 11 November 2026 (Wednesday) | Bielik 11B v3.1 DFlash: a draft model that speeds up response generation |
| 12 November 2026 (Thursday) | Bielik-One: a family of four decision models |
The models will be available on Hugging Face: https://huggingface.co/speakleash
56 languages, with its Polish character intact
We 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, without the model losing its Polish identity.
In practice, you 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 and Asian languages such as Chinese, Japanese and Korean. Caucasian languages have been added too: Armenian, Azerbaijani and Georgian.
The benchmarks show it:
- Belebele (reading comprehension: the model reads a passage and answers questions about it): +8 percentage points,
- INCLUDE (knowledge and reasoning in 44 languages, including regional knowledge): +5 percentage points.
We are not building Bielik just so that Poland has its own language model. We are building it because we believe Polish technology can set standards rather than merely follow them. (…) We want Bielik to become the model of first choice in Europe, and in time beyond it as well.
— Sebastian Kondracki, President of the Board, SpeakLeash Foundation
Bielik doesn’t just answer, it acts
The biggest new feature in version 3.1 is its agentic capabilities. Bielik now combines three mechanisms that work together:
- function calling: it calls external functions and tools,
- structured output: it returns answers in a strictly defined structure,
- reasoning: for harder tasks, it switches to an explicit reasoning mode.
Together, they make Bielik a good fit for agentic systems: systems that don’t stop at generating an answer, but carry out multi-step tasks using the tools available to them.
One of the benchmarks we used to test this is τ²-bench. It simulates realistic customer service scenarios, for example at airlines, in retail and in telecoms. The model has to understand the problem, use the right tools (for example, check a booking or change a service), follow the applicable rules and see the whole case through. A correct answer on its own is not enough. Depending on the scenario, Bielik 3.1 scores 16 to 45 percentage points higher on τ²-bench than version 3.0.
Bielik 3.1 also uses reasoning mode when calling tools. This matters most in more complex tasks, where the model has to plan and carry out a specific sequence of actions and pick the right tools along the way.
— Krzysztof Ociepa, Head of Model Training, SpeakLeash Foundation
Bielik-One: models that don’t write, they decide
On 12 November, we launch the Bielik-One family: four decision models. They are not built to generate text. Instead, they analyse information and assess it against criteria you set in advance. For example, they can assign a ticket to the right category, check whether it meets the requirements, or point to the most likely of the available answers. The models come in different sizes, so you can match them to the task and to the computing power you have.
Where will they shine? Wherever you need a clear-cut classification or assessment rather than a long answer:
- categorising and routing customer requests,
- content moderation,
- initial screening of applications and documents,
- checking compliance with procedures,
- sentiment analysis,
- evaluating answers generated by other AI systems.
Here’s an example: Bielik-One analyses messages in a company inbox and, for each email, gives the probability that it belongs to each category. You can use this distribution to set automation thresholds. The system handles clear-cut cases on its own and passes the rest to a person for review. This lets you automate processes at scale while keeping control over how decisions are made.
In many business processes, we don’t need an elaborate answer from the model, but a specific result, such as assigning a ticket to the right category or a score on a defined scale. Bielik-One was designed precisely for tasks like these.
— Krzysztof Wróbel, Principal AI Researcher, SpeakLeash Foundation
Bielik DFlash: faster, with no loss of quality
On 11 November, we release Bielik 11B v3.1 DFlash, a draft model tailored to the new version of Bielik. It builds on the solution we presented together with ACK Cyfronet AGH in June 2026.
DFlash uses speculative decoding. As a lighter helper model, it proposes the next fragments of a response, which the main model then verifies. This means the main Bielik doesn’t have to generate every next element on its own, so the whole process speeds up without any loss in response quality.
Where can you find the new models?
All the new models, together with their weights, will be released on open science terms as part of the SpeakLeash project, on Hugging Face: https://huggingface.co/speakleash. Check back on each launch day in the schedule above. Download them, test them and tell us what you build with Bielik 3.1.
