On this page
On October 25, 2026 (20:11 Beijing time), Nanbeige announced Nanbeige4.3-33B on X and published weights on Hugging Face: the Instruct model at Nanbeige/Nanbeige4.3-33B and the Base at Nanbeige/Nanbeige4.3-33B-Base. Both list Apache 2.0. Nanbeige is BOSS Zhipin’s LLM research lab, not a whole-company product line launch.
Teacher first, then smaller models
Nanbeige’s stated strategy is to build a high-capacity teacher first, then distill. The Instruct card calls Nanbeige4.3-33B the flagship teacher for the 4.3 series; the X post says it will teach smaller models starting with Nanbeige4.3-3B. That 3B checkpoint is not on Hugging Face at publish time—treat it as a planned follow-on, not a shipped release.
Nanbeige also says the model works with mainstream agent harnesses and is strong at agentic coding and cowork. How harness choice changes cost and outcomes is a theme we cover separately—see Same Model, Different Coding Harness: Pass Rate Barely Moves, the Bill Does.
What Nanbeige claims on benchmarks
Nanbeige’s X post claims state-of-the-art agentic results among open models around the ~30B class. The Instruct model card publishes comparison tables against peers such as Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B; the Base card compares to Gemma4-31B-Base. Those figures are Nanbeige’s own reporting. They have not been retested on this site; for reference only. For a broader view of how coding agents are ranked this month, see Best Coding Agents Right Now, Ranked by Artificial Analysis (October 2026).
How to run it
The Instruct model card documents local serving with Transformers, SGLang, and vLLM. Those recipes have not been run on this site yet. Related reading on another recent open inference stack: DeepSeek Open-Sources DeepInfer, the Inference Framework Behind V4.1 Flash. Feedback on Nanbeige4.3-33B can go through the site-owner contact link in the footer.