Nanbeige4.3-33B open agent model claims lead over Qwen3.8-27B and Gemma4-31B

The Nanbeige team has released Nanbeige4.3-33B, an open-weight agent model it says delivers state-of-the-art agentic performance among open models of about 30 billion parameters [1]. On the team's own tests it leads Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B on most coding and office benchmarks, but not all [1].

What is Nanbeige4.3-33B?

Nanbeige published Nanbeige4.3-33B on Hugging Face under the Apache 2.0 licence with a 512K-token context window. The model card supports English and Chinese and states a maximum context of 524,288 tokens [1].

The team calls it the flagship teacher model for the Nanbeige4.3 series [1]. Its job is to anchor the distillation of agent skills into smaller models, with Nanbeige4.3-3B named as an upcoming example [1].

The card targets complex coding and what it calls professional cowork tasks, and offers settings for reasoning, multi-turn tool use and code-agent workflows [1]. It can be run with Transformers, SGLang or vLLM; the example server commands split the model across four GPUs [1].

How does it score against Qwen, Gemma and Muse?

All models were tested at their highest thinking effort with the same inference settings and harness per benchmark, according to the card [1]. Nanbeige reports 80.9 on Terminal-Bench 2.1 for its model, against 73.0 for Qwen3.8-27B and 47.2 for Gemma4-31B.

The biggest reported gaps are on harder agent tests: 23.8 on Terminal-Bench 4.0 against 6.3 for Qwen3.8-27B, and 65.5 on DeepSWE against 42.5 [1]. On AIME 2026, a maths test, it reports 95.2 against 94.5 for Qwen [1].

Where does Nanbeige4.3-33B fall short?

Qwen3.8-27B still beats Nanbeige4.3-33B on SWE-Bench Pro and GDPval in the team's own comparison table. On SWE-Bench Pro, Qwen scores 61.7 against 60.3, and on GDPval 86.4 against 84.8 [1].

On GPQA Diamond, a science reasoning test, Gemma4-31B leads with 91.3 against 89.5 for Nanbeige [1]. The card's footnotes also show that some tests used GPT-6 Sol as a judge or ran inside Claude Code [1].

Our read

Opinion: A 33B open model that claims to beat its peers by double digits on hard terminal and software-engineering agent tests is worth a test drive, especially under Apache 2.0. But a self-graded scorecard is a pitch, not a result; the real signal will come when independent leaderboards and the promised 3B distillation show whether the agent skills transfer.

What we could not verify

• Every benchmark score comes from the Nanbeige team's own model card; we found no independent evaluation of Nanbeige4.3-33B at publication.
• Whether the comparisons are fully like-for-like; some benchmarks were judged by GPT-6 Sol or run in the Claude Code harness, and GDPval used revised rubrics.
• How much memory or which GPUs the model needs; the card does not state hardware requirements, and claims on social media that it runs on a single consumer GPU or a Mac are unverified.
• When the smaller distilled models such as Nanbeige4.3-3B will be released; the card calls them upcoming without a date.