Polish AI project SpeakLeash and computing center ACC Cyfronet AGH have introduced Bielik-11B-v3.1-Instruct, an 11-billion-parameter text model whose training spans 56 languages, with particular emphasis on Polish and other European languages. The team’s November 10 announcement describes more multilingual training and updates to instruction following, tool use and reasoning.
The model card identifies Apache 2.0 as its license, while the repository also asks users to accept access conditions before obtaining its files. License terms and access requirements are separate details: readers should review both, and check the terms of any derivative or quantized artifact they select.
A regional model with wider reach
The creators say v3.1 received an additional 500 billion training tokens and expands the language count from 32 in v3.0. They report improvements on multilingual benchmarks, but benchmark changes do not guarantee equal performance for every language, dialect or application. AiBrainWorX has not tested the model.
The release also supports a model-specific tool-call format and configurable reasoning. The documentation describes parser plugins for vLLM; a compatible runtime and correct chat template matter if an application expects structured tool calls instead of plain text. Tool arguments should be treated as untrusted input.
What local users should check
The publisher lists companion DFlash draft models and separately distributed quantized versions. A draft model is an accelerator used alongside a compatible target, not a replacement for it; speedups depend on the runtime and workload. The OpenSourceCoder.com model profile links to the original card.
The model card warns that outputs may be incorrect or biased and says the model has no moderation mechanism. Anyone deploying it should evaluate the languages and tasks they actually need, review their safety controls, and measure memory and performance on their own hardware.
Original reporting; primary sources reviewed November 14, 2026. No independent hands-on test claimed.
