SpeakLeash Releases Bielik 11B v3.1 and Decision Models

SpeakLeash and ACK Cyfronet AGH announced in a blog post the staged release of Bielik 11B v3.1, Bielik 11B v3.1 DFlash and the Bielik-One family between 10 and 12 November. The models and their weights are available through the SpeakLeash account on Hugging Face under open science terms.

Bielik 11B v3.1 was trained on another 500 billion tokens and supports 56 languages, up from 32 in version 3.0. It adds function calling, structured output and a reasoning mode for tasks that require several actions and external tools. SpeakLeash reports gains of 8 percentage points on the Belebele reading benchmark and 5 percentage points on the INCLUDE multilingual benchmark.

Bielik 11B v3.1 DFlash uses speculative decoding to propose response segments for verification by the main model, reducing generation time without changing response quality. Bielik-One consists of four models for classification and assessment tasks such as routing requests, moderating content, screening documents, checking procedures and evaluating responses from other AI systems.