Nanbeige4.3-33B: an Apache-2.0 dense 33B model that claims the top agentic scores in the 30B class

The Nanbeige team has released Nanbeige4.3-33B, an open-weight model aimed at coding agents and office-style “cowork” tasks, together with its base checkpoint. Both are under Apache-2.0. The weights went up on Hugging Face on 24 October, and the team announced the model on X on 25 October, after filling in the model card that day.

Nanbeige calls it the “flagship teacher model” of the 4.3 series. The plan in the card is to distill its agent skills into smaller models, naming a coming Nanbeige4.3-3B. Its earlier Nanbeige4 releases on Hugging Face were all 3B models.

What you get

The config describes a plain dense Llama-architecture model: 52 layers, hidden size 7,168, 56 attention heads with 8 key-value heads, and a 166,912-token vocabulary. Hugging Face counts 33.1 billion parameters in BF16, which works out to about 66 GB of weights across 14 files. The card gives a context length of 524,288 tokens (512K).

It runs with Transformers (with trust_remote_code), SGLang or vLLM. The card’s server examples use tensor parallelism across four GPUs and the Qwen3 reasoning and Qwen3-Coder tool-call parsers. Thinking is on by default. A preserve_thinking switch keeps earlier reasoning in multi-turn chats, and Nanbeige recommends turning it on for tool use, office tasks and code agents. XML is the recommended tool-call format, with JSON also supported. Suggested sampling is temperature 1.0 for agent work and 0.6 for reasoning and chat.

The benchmark table

Nanbeige compares the model with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all “at their highest thinking effort” with the same harness and settings. Selected rows, all from the card:

Benchmark	Nanbeige4.3-33B	Qwen3.8-27B	Gemma4-31B
Terminal-Bench 2.1	80.9	73.0	47.2
Terminal-Bench 4.0	23.8	6.3	0.0
DeepSWE	65.5	42.5	2.7
SWE-Bench Pro	60.3	61.7	51.2
GDPval	84.8	86.4	71.4
Automation Bench	54.2	38.3	9.7
GPQA Diamond	89.5	89.2	91.3

Qwen3.8-27B wins SWE-Bench Pro and GDPval, and Gemma4-31B wins GPQA Diamond. Nanbeige leads the other rows. The footnotes matter: Nanbeige was run with preserve_thinking=true, Terminal-Bench 4.0, DeepSWE and SWE-Bench Pro used mini-swe-agent v2.4.6, NL2Repo and JobBench used the Claude Code harness, and GDPval used “revised rubrics”. The card also reports results across 35 occupations on JobBench.

What we could not check

There is no paper or technical report yet, so the card says nothing about training data, training compute or how the teacher-first recipe works. No outside group has published results, and the Nanbeige account has no official quantized builds of the 33B model so far, though community GGUF and MLX conversions from other accounts appeared on Hugging Face within a day. The card lists a contact address at kanzhun.com, the domain of Kanzhun, which runs the BOSS Zhipin recruitment app.

At 33 billion dense parameters, every token runs through the whole model, so it needs more compute per token than mixture-of-experts models of similar total size. If the agent scores hold up in independent runs, it would be a strong open option for teams who want a dense model in this class.