Nanbeige4.3-33B

Nanbeige4.3-33B is an open-weight language model released by the Nanbeige team under an Apache 2.0 license, developed specifically for agentic coding, reasoning, and professional collaborative workflows. Operating as a dense 33-billion-parameter model, it serves as the flagship teacher model in the Nanbeige4.3 series, laying the technical foundation for future smaller model distillation pipelines.

The model is built on a conventional Llama-derived architecture featuring 52 layers, a hidden size of 7,168, and grouped-query attention (GQA) with 56 query heads and 8 key/value heads. It supports a large context window of up to 524,288 tokens with text-only input and output modalities. The chat template natively accommodates reasoning blocks and tool-use scenarios through configurable parameters like enable_thinking and preserve_thinking.

Capabilities and Performance

Nanbeige4.3-33B excels at complex software engineering, automated tool use, and multi-turn collaborative tasks. It delivers strong performance across various domain-specific benchmarks, showing particular strength in terminal use, coding tasks, and logic-heavy workflows. Its design emphasizes structured reasoning outputs and robust integration within multi-step agent environments.