Nanbeige4.3-33B

Image: Nanbeige Logo

1. Introduction#

In Nanbeige4.3, we take a foundational step toward pushing the boundaries of lightweight agents. Our strategy centers on building a high-capacity teacher agent first, which will subsequently anchor the distillation of advanced agentic capabilities into smaller models.

To this end, we first introduce Nanbeige4.3-33B, designed as the flagship teacher model for the Nanbeige4.3 series. Delivering state-of-the-art agentic performance among open models in the ~30B parameter class, Nanbeige4.3-33B excels at complex coding and professional cowork tasks. It not only offers a powerful, cost-effective option for immediate real-world deployment, but also lays the necessary foundation to power our upcoming small model distillation pipeline (e.g., for Nanbeige4.3-3B).

Image: image

2. Model Performance#

We compare Nanbeige4.3-33B with recently released frontier models at around the 30B-parameter scale: Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B. All models are evaluated at their highest thinking effort. For each benchmark, we use the same inference hyperparameters and harness across models.

| Category | Benchmark | Nanbeige4.3-33B1 | Qwen3.8-27B | Muse-Glimmer-30B | Gemma4-31B

| Agentic Coding | Terminal-Bench 2.12 | 80.9 | 73.0 | 61.8 | 47.2

| Terminal-Bench 4.03 | 23.8 | 6.3 | 1.6 | 0.0

| DeepSWE3 | 65.5 | 42.5 | 4.4 | 2.7

| SWE-Bench Pro3 | 60.3 | 61.7 | 51.2 | 36.9

| SWE-Atlas-QnA4 | 50.8 | 41.9 | 17.7 | 3.3

| NL2Repo5 | 53.0 | 48.3 | 27.4 | 33.8

| Scicode | 52.6 | 46.7 | 44.9 | 45.5

| SkillsBench | 61.1 | 56.4 | 39.7 | 28.9

| Agentic Cowork | Automation Bench6 | 54.2 | 38.3 | 15.2 | 9.7

| JobBench7 | 44.8 | 34.5 | 16.6 | 9.0

| GDPval8 | 84.8 | 86.4 | 75.8 | 71.4

| Claw-Evalpass^39 | 78.3 | 75.2 | 57.3 | 44.2

| Toolathlon-Verified10 | 70.4 | 68.5 | 47.2 | 27.8

| DeepPlanning12 | 63.3 | 49.2 | 21.4 | 19.4

| EnterpriseOps-Gym12 | 41.0 | 40.2 | 30.5 | 27.3

| YC-bench11 | $1,025,437 | $641,290 | $513,978 | $265,152

| Tau3-Bench12 | 48.8 | 48.5 | 33.7 | 18.6

| Reasoning | GPQA Diamond | 89.5 | 89.2 | 83.5 | 91.3

| AIME 2026 | 95.2 | 94.5 | 94.7 | 89.2

| AA-LCR | 84.0 | 82.0 | 83.3 | 72.0

1 Nanbeige4.3-33B is evaluated with preserve_thinking=true in the chat template.

2 Terminal-Bench 2.1 is evaluated using Terminus 2.

3 Terminal-Bench 4 and DeepSWE are evaluated using mini-swe-agent v2.4.6.

4 SWE-Atlas-QnA is evaluated using mini-swe-agent v2.4.6. Answers are evaluated against 1,304 task-specific rubrics using GPT-6 Sol as the judge, with medium reasoning effort.

5 NL2Repo is evaluated using Claude Code.

6 Automation Bench v1.0.6 includes 600 publicly released tasks.

7 JobBench is evaluated with the Claude Code harness equipped with a search tool.

8 GDPval is scored using revised rubrics.

9 Claw-Eval is evaluated exclusively on the Claw-Eval General subset (comprising 157 single-turn user interaction tasks).

10 Toolathlon-Verified sets no max_tokens limit for individual model requests.

11 YC-bench retains only the most recent 10 turns of user dialogue as historical context before each model call.

12 DeepPlanning, EnterpriseOps-Gym, Tau3-Bench are evaluated using their official repositories.

Nanbeige4.3-33B leads on most of the benchmarks, showing strong performance in both agentic coding and cowork tasks. Beyond its breadth across benchmarks, Nanbeige4.3-33B shows strong performance on agent tasks across a wide range of professions and industries. For example, on JobBench, it outperforms the baseline in most of the 35 occupations evaluated.

Image: occupation

3. Quickstart#

The tokenizer provides a configurable chat template for reasoning and tool-use scenarios:

• enable_thinking controls whether the model generates reasoning for the current response. It is enabled by default; set it to False for non-thinking mode.
• preserve_thinking controls whether reasoning from previous assistant turns is retained in a multi-turn conversation. We recommend False for general chat and question answering, and True for multi-turn tool use, office tasks, and code-agent workflows.
• Passing tools enables the tool-use template. We recommend tool_call_format="xml" for the best tool-calling performance; json is also supported for compatibility.

The model supports a context length of up to 524,288 tokens (512K).

We recommend adjusting the inference settings according to the target scenario:

You can run the model with Transformers, SGLang, or vLLM:

Transformers#

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = 'Nanbeige/Nanbeige4.3-33B'
tokenizer = 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
)

messages = [{'role': 'user', 'content': 'Which number is bigger, 9.11 or 9.8?'}]
inputs = tokenizer.apply_chat_template(
  messages,
  add_generation_prompt=True,
  return_tensors='pt'
).to(model.device)
output_ids = model.generate(inputs, max_new_tokens=1024)
response = tokenizer.decode(output_ids[0][inputs.shape[-1]:], skip_special_tokens=True)
print(response)

SGLang#

MODEL_PATH=/path/to/your/Nanbeige4.3-33B
python -m sglang.launch_server \
  --model-path ${MODEL_PATH} \
  --host 0.0.0.0 \
  --port 8000 \
  --tp-size 4 \
  --mem-fraction-static 0.8 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

vLLM#

MODEL_PATH=/path/to/your/Nanbeige4.3-33B
vllm serve ${MODEL_PATH} \
  --host 0.0.0.0 \
  --port 8000 \
  --tensor-parallel-size 4 \
  --gpu-memory-utilization 0.8 \
  --enable-auto-tool-choice \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

4. Contact#

If you have any questions, please open an issue in this repository or contact us at nanbeige@kanzhun.com.

We also welcome collaborations to build domain-specific and industry-focused agents, including joint work on data and training.

5. Citation#

If you find our model useful or would like to use it in your own work, please cite this Hugging Face project:

@misc{nanbeige43chat2026,
  title        = {Nanbeige4.3-33B},
  author       = {Nanbeige Team},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Nanbeige/Nanbeige4.3-33B}}
}