Nanbeige4.3-33B: an Apache 2.0 open agent model, vendor-benchmarked

Nanbeige, a model team that lists a kanzhun.com contact address, released Nanbeige4.3-33B as open weights under the Apache 2.0 licence, according to its Hugging Face model card. The card describes it as the flagship teacher model of the Nanbeige4.3 series: a 33-billion-parameter model aimed at agentic coding and what it calls professional cowork tasks, with a context window of up to 524,288 tokens. The repository was created on 24 October 2026.

What Nanbeige reports

Nanbeige compared the model with three recent models of about the same size: Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all at their highest thinking effort, with the same inference settings and harness for each. It reports the best score on most benchmarks. These are Nanbeige's numbers, not an independent evaluation.

• Agentic coding: Terminal-Bench 2.1 80.9 (Qwen3.8-27B 73.0); Terminal-Bench 4.0 23.8 (6.3); DeepSWE 65.5 (42.5). On SWE-Bench Pro it scores 60.3 against 61.7 for Qwen3.8-27B.
• Office-style agent work: Automation Bench 54.2 (38.3); JobBench 44.8 (34.5); Toolathlon-Verified 70.4 (68.5); EnterpriseOps-Gym 41.0 (40.2). On GDPval it scores 84.8 against 86.4 for Qwen3.8-27B.
• Reasoning: AIME 2026 95.2 (94.5); GPQA Diamond 89.5, behind Gemma4-31B at 91.3.
• Nanbeige also says that on JobBench, which covers 35 occupations, the model beats the baseline in most of them.

Read the footnotes before you trust the table

The card's footnotes show the results depend on setup. Terminal-Bench 2.1 was run with the Terminus 2 harness, while Terminal-Bench 4.0 and DeepSWE used mini-swe-agent v2.4.6. The model was evaluated with preserve_thinking turned on, which keeps earlier reasoning across turns. SWE-Atlas-QnA answers were graded by another model, GPT-6 Sol, against 1,304 rubrics, and GDPval was scored with revised rubrics. Change the harness or the grader and the gaps may move.

What it takes to run, as far as the card says

The card gives Transformers, SGLang and vLLM recipes. Both server examples use tensor parallelism across 4 GPUs (tp-size 4) and reserve 80% of GPU memory, with the Qwen3 reasoning parser and tool-call parser. The weights are stored in BF16. The card does not state a minimum GPU, a latency figure or a cost per task, so those are unknown until someone measures them.

For agent use Nanbeige recommends temperature 1.0, top_p 0.95, up to 65,536 new tokens, XML tool-call format and preserve_thinking set to true. Its Transformers example loads the model with trust_remote_code=True, which runs code from the repository; pin a specific revision and review that code before you do the same.

What is missing

There is no closed-model comparison, so the card does not say how this model compares with the hosted model you may use today. There is no safety or prompt-injection evaluation, no independent run, and no evidence yet from a real deployment. Nanbeige also says its aim is to distil this model into smaller ones, such as a Nanbeige4.3-3B, which have not been released.

What to do

• If you self-host an open-weight agent model, add this one to your next bake-off using your own tasks and your own harness, and compare cost per completed task, not benchmark score.
• Run it under the settings Nanbeige recommends before judging it; the scores assume preserve_thinking and the XML tool-call format.
• Keep tool permissions and sandboxing the same as for any other model: a strong benchmark score says nothing about how it handles hostile input.