Nanbeige4.3-33B

Nanbeige4.3-33B is a new large language model open-sourced on Hugging Face on October 24, 2026 by Nanbeige, the LLM team at Kanzhun (the company behind BOSS Zhipin). It is the flagship "teacher" model of the Nanbeige4.3 series.

The weights are released under Apache-2.0 (alongside a base model, Nanbeige4.3-33B-Base). The model targets agentic coding and agentic cowork (professional office tasks); the team says it delivers state-of-the-art agentic performance among open models around the 30B scale and will anchor distillation into smaller models such as Nanbeige4.3-3B.

On October 25, AI commentator @kimmonismus wrote on X that it beats "similarly sized Qwen 3.8 27b and Gemma models on most benchmarks" (300+ likes), X News ran a story titled "Nanbeige4.3-33B Tops Agentic Coding Benchmarks in 30B Class", and the model entered Hugging Face's trending list.

Key specs

Item	Detail
Size	~33B (llama architecture, trust_remote_code)
Languages	English, Chinese
Context length	up to 524,288 tokens (512K)
Reasoning	thinking on by default (enable_thinking), can be turned off; preserve_thinking keeps reasoning from earlier turns
Tool calling	tool_call_format="xml" recommended, json also supported
License	Apache-2.0

Reported benchmarks (vs Qwen3.8-27B / Muse-Glimmer-30B / Gemma4-31B, all at highest thinking effort)

Category	Benchmark	Nanbeige4.3-33B	Qwen3.8-27B
Agentic coding	Terminal-Bench 2.1	80.9	73.0
Agentic coding	DeepSWE	65.5	42.5
Agentic coding	SWE-Bench Pro	60.3	61.7
Agentic coding	NL2Repo	53.0	48.3
Agentic cowork	Automation Bench	54.2	38.3
Agentic cowork	Claw-Eval (pass^3)	78.3	75.2
Agentic cowork	GDPval	84.8	86.4
Reasoning	GPQA Diamond	89.5	89.2
Reasoning	AIME 2026	95.2	94.5

The team also reports that on JobBench, which covers 35 occupations, it beats the baseline in most of them. All figures are self-reported; harness details are in the model card footnotes.

Positioning

Nanbeige4.3's strategy is to build a high-capacity teacher agent first, then distil its agentic skills into smaller models. The 33B model is both deployable today and the teacher for upcoming small models like a 3B version.

How to use it

• Weights: Hugging Face Nanbeige/Nanbeige4.3-33B (base: Nanbeige/Nanbeige4.3-33B-Base)
• Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = 'Nanbeige/Nanbeige4.3-33B'
tok = AutoTokenizer.from_pretrained(model_id, use_fast=False, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype='auto', device_map='auto', trust_remote_code=True)

• vLLM (the official example uses 4-way tensor parallelism):

vllm serve Nanbeige/Nanbeige4.3-33B --tensor-parallel-size 4 \
  --enable-auto-tool-choice --reasoning-parser qwen3 --tool-call-parser qwen3_coder

• SGLang: --reasoning-parser qwen3 --tool-call-parser qwen3_coder

Recommended settings

Scenario	Temperature	Top_p	Max new tokens
Agentic / tool use	1.0	0.95	65,536
Reasoning / chat	0.6	0.95	262,144

Use preserve_thinking=True for multi-turn tool use, office tasks and code agents; keep it off for general chat.

Who it's for

• Teams that want a 30B-class open coding/agent model on their own GPUs
• Developers who need very long context (512K) and English–Chinese ability

Contact: nanbeige@kanzhun.com; see the Hugging Face model card for details.