Nanbeige has released Nanbeige4.3-33B, a roughly 33-billion-parameter open-weight model it calls the flagship teacher for its Nanbeige4.3 series, according to @nanbeige. The headline claim is state-of-the-art agentic performance among open models of about 30B parameters. Those comparisons are Nanbeige's own.

What's new

The model page says Nanbeige built this model first so it can distil agentic skills into smaller models, starting with Nanbeige4.3-3B. A Base variant, Nanbeige4.3-33B-Base, is posted alongside it on Hugging Face. The weights are listed in BF16, and the context window is up to 524,288 tokens (512K). The page gives the parameter count but not the architecture, such as whether the model is dense or mixture-of-experts.

Benchmarks against ~30B rivals

Nanbeige compares the model with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, each run at its highest thinking effort. A selection of the rows:

Benchmark

Nanbeige4.3-33B

Qwen3.8-27B

Muse-Glimmer-30B

Gemma4-31B

Terminal-Bench 2.1

80.9

73.0

61.8

47.2

DeepSWE

65.5

42.5

4.4

2.7

SWE-Bench Pro

60.3

61.7

36.9

51.2

Automation Bench

54.2

38.3

15.2

9.7

GDPval

84.8

86.4

75.8

71.4

GPQA Diamond

89.5

89.2

83.5

91.3

AIME 2026

95.2

94.5

94.7

89.2

Nanbeige's scores were run with preserve_thinking=true in the chat template; the page also lists per-benchmark harness and subset notes.

The model leads on Terminal-Bench 2.1, DeepSWE and Automation Bench. It trails Qwen3.8-27B on SWE-Bench Pro and GDPval, and Gemma4-31B on GPQA Diamond. The benchmarks use different harnesses, including Terminus 2, mini-swe-agent v2.4.6 and Claude Code, and SWE-Atlas-QnA answers are judged by GPT-6 Sol. Those differences make cross-row comparisons less direct.

Running it and what to check

Nanbeige recommends temperature 1.0 and top_p 0.95 with up to 65,536 new tokens for agentic and tool-use tasks. For reasoning and chat, it recommends temperature 0.6 and top_p 0.95 with up to 262,144 new tokens. The model can be served with Transformers, SGLang or vLLM. The SGLang and vLLM examples use tensor parallelism of 4, a qwen3_coder tool-call parser, and a qwen3 reasoning parser. For tool use, the page recommends the XML tool-call format.

Nanbeige does not list a hosted API or a per-token price. It describes inference as cost-efficient but gives no figures. The model page does not list a licence, so check the repository before using the weights commercially.

The Base model is positioned as a foundation for post-training. Against Gemma4-31B-Base, the authors report 66.9 on MMLU-Pro (versus 59.3) and 62.0 on GPQA (versus 41.5). For other recent open-weight releases at a similar scale, see our report on Ornith-1.5-122B.

Who should care

Teams running open models in agent loops on their own hardware should test the model on their own tasks before switching. Teams planning to fine-tune or distil should look at the Base model. The timing for Nanbeige4.3-3B, the first smaller model in the series, has not been announced.

Sources