Nanbeige4.3-33B: an open 33B agent model

What launched?

Over the weekend the Nanbeige team published Nanbeige4.3-33B on Hugging Face (model card). It is the first model in the 4.3 series and serves as a teacher: the team plans to distill its agentic skills into smaller models later. It is usable on its own today, with weights available under Apache 2.0.

• Dense model with about 33 billion parameters, for English and Chinese.
• Context: up to 524,288 tokens (512K).
• Reasoning can be toggled. For multi-turn tool use and code agents, reasoning from earlier turns can be kept (preserve_thinking).
• Runs on Transformers, SGLang or vLLM. The sample configs use 4-way tensor parallelism, and the card mentions no hosted API.

What the benchmarks show

The team compared it with three models around 30B: Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all at their highest thinking effort with the same harness. It scored 80.9 on Terminal-Bench 2.1 (Qwen3.8-27B: 73.0) and 23.8 on the harder Terminal-Bench 4.0 (Qwen: 6.3). On DeepSWE it reached 65.5 versus 42.5, and 95.2 on AIME 2026.

It does not win everywhere. Qwen3.8-27B is ahead on SWE-Bench Pro (61.7 vs. 60.3) and on GDPval office tasks (86.4 vs. 84.8). On GPQA Diamond, Gemma4-31B leads with 91.3 against Nanbeige's 89.5. All numbers come from the developers, with no independent verification yet.

Source: Nanbeige. JobBench results across 35 occupations: Nanbeige4.3-33B (44.8), Qwen3.8-27B (34.5) and Muse-Glimmer-30B (16.6).

Why it matters

Open models around 30B matter because they can run on your own infrastructure without sending data out. If the reported agentic results hold up in independent tests, Nanbeige4.3-33B is a serious option for internal code agents and office automation. Benchmark it against Qwen3.8-27B on your own tasks, since the gap between the two varies by task type. The promised smaller distilled versions are worth watching too, as they may fit on a single GPU.