saved
Nanbeige4.3-33B
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
Nanbeige, BOSS Zhipin's LLM lab, released Nanbeige4.3-33B, an Apache 2.0 dense open model it calls the best ~30B model for agents. It is built as a teacher for smaller models, starting with Nanbeige4.3-3B. Its model card beats Qwen3.8-27B on most agentic coding and cowork benchmarks, including 80.9 against 73.0 on Terminal-Bench 2.1.
ideas
- The big model exists to make small ones. Nanbeige calls 4.3-33B the flagship teacher of the series; the model card says it lays the foundation for a distillation pipeline, starting with Nanbeige4.3-3B.
- Agentic coding is where it leads. Against Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B, all at highest thinking effort on shared harnesses, it scores 80.9 on Terminal-Bench 2.1 (Qwen3.8-27B 73.0) and 65.5 on DeepSWE (42.5).
- It does not win every row. Qwen3.8-27B keeps SWE-Bench Pro (61.7 against 60.3) and GDPval (86.4 against 84.8), and Gemma4-31B tops GPQA Diamond at 91.3 against 89.5.
- It is tuned for long agent sessions. The card lists 512K tokens of context, a preserve_thinking flag that keeps earlier reasoning for multi-turn tool use and code agents, and an XML tool-call format it recommends over JSON.
- The weights target servers; local builds are community-made. Nanbeige ships 33.1B-parameter safetensors with Transformers, SGLang, and vLLM recipes using four-way tensor parallelism, while GGUF and MLX quantizations came from community uploaders within a day.
quotes
“To push the performance ceiling of small agents, our strategy in Nanbeige4.3 starts with building a powerful teacher first.”
“When integrated into mainstream agent harnesses, it tackles real-world tasks with cost-efficient inference.”
“It will serve as a teacher for upcoming smaller Nanbeige models, starting with Nanbeige4.3-3B.”