SpeakLeash Releases Bielik-11B-v3.1-Instruct, a Multilingual LLM Trained on 500B Tokens

·2026.11.11 01:30

Key point

The 11B-parameter model was trained on 500B tokens and uses modified GRPO reinforcement learning across 14 environments.

SpeakLeash and ACC Cyfronet AGH released Bielik-11B-v3.1-Instruct, an 11B-parameter causal decoder-only model under the Apache 2.0 license.
The model supports 56 languages with a specific optimization for European languages and Polish, having been trained on 500B tokens.
Training involved three stages: pretraining, supervised fine-tuning on 26M samples, and reinforcement learning using modified GRPO across 14 diverse environments.
Development utilized the Athena and Helios supercomputers within the PLGrid environment and the Gaia AI Factory, funded by the EuroHPC Joint Undertaking.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.