Nanbeige LLM Lab · 2026-10-24 · major
Nanbeige4.3-33B — open 33B agent model leads its class on Terminal-Bench
Nanbeige4.3-33B is an Apache-2.0 open-weight model for agentic coding and office work, with a 512K context. It scores 80.9 on Terminal-Bench 2.1 versus 73.0 for Qwen3.8-27B, in Nanbeige's own tests.
A 33B open-weight model tuned for coding agents and office-style agent tasks, with a 512K context.
Key specs
| Terminal bench 2.1 | 80.9 |
|---|---|
| Deep swe | 65.5 |
Quick facts
| Maker | Nanbeige LLM Lab (Kanzhun) |
|---|---|
| Parameters | 33B |
| Context window | 512K tokens |
| License | Apache-2.0 |
| Checkpoints | Instruct + Base |
| Runs on | Transformers, SGLang, vLLM |
Benchmarks
What is it?
Nanbeige4.3-33B is the first model in the Nanbeige4.3 series from the Nanbeige LLM Lab at Kanzhun, the company behind BOSS Zhipin. The team calls it the series' teacher model: a strong mid-size model that later, smaller models will be distilled from. It targets two jobs, agentic coding and professional 'cowork' tasks such as office workflows.
How does it work?
The model uses a thinking mode that is on by default and can be switched off in the chat template. A 'preserve_thinking' option keeps earlier reasoning across turns, which the authors recommend for tool use and code agents. Tool calls are formatted from a passed tool list, with XML recommended. The card suggests temperature 1.0 for agent work and 0.6 for reasoning and chat.
Why does it matter?
A model around 30B parameters fits on a single multi-GPU server, so teams can run a capable coding agent on their own hardware under a permissive license. On the authors' table it leads Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B on most agentic coding and cowork benchmarks. These scores come from Nanbeige's own harnesses and are not yet independently checked.
Who is it for?
teams self-hosting coding agents
Frequently asked questions
- Is Nanbeige4.3-33B open source?
- Nanbeige4.3-33B is released as open weights on Hugging Face under the Apache-2.0 license, which allows commercial use. The Nanbeige team published two checkpoints: the instruction-tuned Nanbeige4.3-33B and the pretrained Nanbeige4.3-33B-Base, both in BF16 safetensors format.
- How does Nanbeige4.3-33B compare to Qwen3.8-27B?
- On Nanbeige's own benchmark table, Nanbeige4.3-33B beats Qwen3.8-27B on Terminal-Bench 2.1 (80.9 vs 73.0), DeepSWE (65.5 vs 42.5) and AIME 2026 (95.2 vs 94.5). Qwen3.8-27B stays ahead on SWE-Bench Pro (61.7 vs 60.3) and GDPval (86.4 vs 84.8). All models were run at their highest thinking effort.
- What hardware do I need to run Nanbeige4.3-33B?
- The Nanbeige4.3-33B model card shows serving examples for SGLang and vLLM that split the model across four GPUs with tensor parallelism, using the qwen3 reasoning parser. The model also loads in Hugging Face Transformers with trust_remote_code enabled. The card does not publish a minimum GPU memory figure.
- Will there be a smaller Nanbeige4.3 model?
- Yes, according to the Nanbeige team. They describe Nanbeige4.3-33B as the flagship teacher of the series, built to anchor distillation of agentic skills into smaller models, and name a planned Nanbeige4.3-3B. That smaller model is not released yet, and no release date has been given.
Try it
vllm serve Nanbeige/Nanbeige4.3-33B --trust-remote-code --tensor-parallel-size 4