Nanbeige Opens 33B Coder With 512K Context

Nanbeige LLM Lab, an AI team that MarkTechPost reports sits inside Chinese recruitment platform BOSS Zhipin, has released Nanbeige4.3-33B, an open-weight coding and agent model, under the Apache 2.0 license. The lab says the 33-billion-parameter model beats Alibaba's Qwen3.8-27B, Google's Gemma4-31B and Meta's Muse-Glimmer-30B on 17 of 20 benchmarks in its own tests.

Key takeaways

• Nanbeige4.3-33B is a dense 33B model with a 524,288-token context window, released with a separate base checkpoint under Apache 2.0.
• On the lab's own table it scores 80.9 on Terminal-Bench 2.1 versus 73.0 for Qwen3.8-27B, but trails on SWE-Bench Pro, GDPval and GPQA Diamond.
• All scores are self-reported, and MarkTechPost found some competitor figures that conflict with those labs' own model cards.

What is Nanbeige4.3-33B?

The Hugging Face model card calls it the "flagship teacher model" of the Nanbeige4.3 series. The lab plans to distill it into smaller models, starting with a Nanbeige4.3-3B. It supports English and Chinese, and the card lists a kanzhun.com contact address.

The model's configuration file shows a standard Llama-style design: 52 layers, a hidden size of 7,168, and 56 query heads sharing 8 key-value heads. The vocabulary holds 166,912 tokens. It is a dense model, not a mixture of experts, so all 33 billion parameters run on every token.

Nanbeige4.3-33B benchmarks vs. Qwen and Gemma

The lab compared its model with three open models of similar size, all run at their highest thinking effort with the same settings. The gaps are widest on long agentic coding tasks.

The harder Terminal-Bench 4 shows how far 30B-class models still have to go. Nanbeige scored 23.8, against 6.3 for Qwen3.8-27B and 0.0 for Gemma4-31B. On the AIME 2026 math test it posted 95.2.

The base model card makes similar claims before fine-tuning. Nanbeige4.3-33B-Base scored 66.9 on MMLU-Pro, compared with 59.3 for Gemma4-31B-Base, and 86.8 versus 76.9 on the Chinese-language CMMLU test.

Caveats on the numbers

No outside group has reproduced the results. MarkTechPost reported that some rival figures in Nanbeige's table do not match the other labs' own cards. It said Nanbeige lists Muse-Glimmer-30B at 61.8 on Terminal-Bench 2.1, while Meta reports 51.7, and called those rows provisional.

Running the model locally takes serious hardware. MarkTechPost estimated about 66 GB for the BF16 weights alone. The card provides recipes for Transformers, SGLang and vLLM, and community quantized builds for llama.cpp, Ollama and LM Studio are listed on the hub.

Context: the mid-size open model race

The release adds a new name to a crowded field of Chinese open-weight models aimed at coding agents. Qwen's larger models have already pressed U.S. labs on agentic tests, as when Qwen3.8 Max climbed above every American lab but two on one independent index. Nanbeige is betting that a 33B model under a permissive license can win on cost of deployment.

What's next

The next step is the promised Nanbeige4.3-3B distilled model. Independent runs on Terminal-Bench and SWE-Bench will show whether the lab's lead holds. At MarkTechPost's publication time, no hosted inference provider offered the model.

FAQ

Is Nanbeige4.3-33B free for commercial use?

Yes. Both the instruct model and the base checkpoint are published under Apache 2.0, which permits commercial use, modification and redistribution.

How long is Nanbeige4.3-33B's context window?

The model accepts up to 524,288 tokens, often described as 512K. For agent tasks the lab recommends up to 65,536 new tokens per response.

Are the benchmark results verified?

No. All scores come from Nanbeige's own tests and have not been independently reproduced.

References

• Nanbeige/Nanbeige4.3-33B — Hugging Face (October 2026)
• Nanbeige/Nanbeige4.3-33B-Base — Hugging Face (October 2026)
• Nanbeige Releases Nanbeige4.3-33B, a 33B Open Agentic Coding Model — MarkTechPost (Oct. 25, 2026)