Built as a teacher model, it leads most of its class

The Nanbeige team has published Nanbeige4.3-33B on Hugging Face under the Apache 2.0 licence, together with a base model. It is the flagship of the Nanbeige4.3 series and is framed as a teacher: the plan in the model card is to build a high-capacity agent first and then distil its agentic abilities into smaller models. The card positions it for complex coding and what it calls professional cowork tasks, with a context length of up to 524,288 tokens.

The comparison is with three models of similar size, Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all run at their highest thinking effort with the same inference settings and harness. The widest gaps are in agentic coding. On Terminal-Bench 4.0 Nanbeige scores 23.8, against 6.3 for Qwen3.8-27B and 0.0 for Gemma4-31B; on DeepSWE it scores 65.5 against 42.5 for Qwen; on Terminal-Bench 2.1, 80.9 against 73.0. Its YC-bench result, a simulated business run scored in dollars, is $1,025,437 against $641,290 for Qwen.

It does not lead everywhere, and the card's own table shows where. Qwen3.8-27B is ahead on SWE-Bench Pro, 61.7 to 60.3, and on GDPval, 86.4 to 84.8, while Gemma4-31B leads GPQA Diamond, 91.3 to 89.5. Several results depend on choices set out in the footnotes: Nanbeige is evaluated with earlier reasoning preserved in the chat template, one benchmark is judged by GPT-6 Sol, and different benchmarks run under different agent harnesses, including Claude Code. Every number is Nanbeige's own.

For deployment the card covers Transformers, SGLang and vLLM, with example configurations using tensor parallelism across four GPUs. It recommends keeping earlier reasoning for multi-turn tool use and code-agent work and dropping it for ordinary chat, with a temperature of 1.0 for agentic tasks and 0.6 for reasoning and chat. The smaller distilled models the card points to are not yet on Hugging Face.