11.17B (11,168,796,672; dense)
32,768 tokens
Apache 2.0
Open-weight Polish-first multilingual instruction model expanding the Bielik 11B v3 family from 32 to 56 supported languages, including every official language of the European Union as well as Arabic, Chinese, Japanese, Korean, Armenian, Azerbaijani, Georgian, Hebrew, Hindi, Indonesian, Persian, Vietnamese, and other regional languages. Bielik-11B-v3.1-Instruct is a dense causal decoder with 50 transformer layers, a 4,096-dimensional hidden state, 14,336-dimensional feed-forward layers, 32 attention heads, eight key-value heads, grouped-query attention, a 32,128-token vocabulary, and a 32,768-token context window. The official model card reports a three-stage training pipeline comprising an additional 500 billion tokens of multilingual continued pretraining, supervised fine-tuning on 26 million samples, and reinforcement learning using modified GRPO across 14 environments. The model was trained using Polish HPC resources including the Athena and Helios supercomputers and the Gaia AI Factory. Version 3.1 adds agent-oriented behavior through OpenAI-style function calling, structured JSON output, multi-step tool use, and explicit reasoning. Its chat template supports reasoning_effort levels none, low, medium, high, and max, along with automatic, required, disabled, and named tool choices. SpeakLeash reports improvements over Bielik 11B v3.0 of eight percentage points on Belebele, five points on INCLUDE, and between 16 and 45 points on τ²-bench depending on the customer-service scenario. The principal BF16 checkpoint is approximately 22.34 GB and uses a Hugging Face click-through access gate. Official public conversions include GGUF in Q4_K_M, Q5_K_M, Q6_K, Q8_0, and BF16 formats, plus AWQ, FP8-Dynamic, NVFP4, NVFP4A16, MLX 4-bit, and MLX 8-bit checkpoints. SpeakLeash also provides its Bielik-specific vLLM tool-call and reasoning parsers. The model is released under Apache 2.0.