All stories
Nanbeige4.3-33B Launches Community reports

Nanbeige4.3-33B aims to teach smaller agents, with weights available now

The model card positions the 33B checkpoint as a teacher for future distillation and reports strong, but mixed, agent benchmarks.

Sources (2)

Why now October 24 repository creation; October 25 update. Backfilled October 27 after a publication gap; this site's publication time is not the event date.

Nanbeige4.3-33B is a downloadable 33-billion-parameter model positioned as a teacher for smaller agents. Its Hugging Face repository was created October 24 and updated October 25. The English model card describes coding and office-style agent tasks and says a smaller distilled model is planned. That future model should not be confused with the checkpoint available today.

The card declares an Apache 2.0 license and a 524,288-token context window. It supplies Transformers, SGLang and vLLM examples, along with thinking-mode settings. For multi-turn tool use, the authors recommend preserving reasoning history; ordinary chat has different suggested settings. Runtime and hardware requirements still need to be checked against the chosen deployment configuration.

The authors report competitive benchmark results, including 80.9 on Terminal-Bench 2.1. Their own table also shows the model behind Qwen3.8-27B on SWE-Bench Pro and GDPval. The useful reading is a promising candidate with a mixed evaluation profile, not a universal winner.

For local-agent builders, the release offers weights and concrete inference guidance to inspect. Its teacher-model rationale and benchmark claims remain the developers’ account. AITrending has not reproduced those scores or verified the effectiveness of the planned distillation.