~/AI/speakleash-releases-bielik-11b-v3-1-instruct-multilingual-language-model

SpeakLeash Releases Bielik-11B-v3.1-Instruct Multilingual Language Model

SpeakLeash has released Bielik-11B-v3.1-Instruct, an 11-billion parameter generative text model optimized for instruction following, tool use, and reasoning with a strong focus on Polish and European languages. This release provides a high-performing, open-weight regional model that advances linguistic capabilities for European languages outside of dominant English-centric models, leveraging public supercomputing infrastructure. The training pipeline included pretraining on 500 billion tokens across 56 languages, SFT on 26 million samples, and reinforcement learning using modified GRPO across 14 diverse environments via the ALLaMo framework.

## BACKGROUND

SpeakLeash is an open-source project focused on gathering Polish language data and developing NLP tools tailored to the Polish linguistic ecosystem. The model was trained using Polish high-performance computing resources like the Athena and Helios supercomputers within the PLGrid environment.

## REFERENCES

## KEYWORDS

#AI#LLM#Open Source#Multilingual#Polish

$ subscribe --daily