BOSS Zhipin’s AI Lab Releases Nanbeige4.3-33B, an Open-Weight Model With a 512K Context Window
Nanbeige4.3-33B is a 33-billion-parameter open-weight model from the AI lab of Chinese recruitment platform BOSS Zhipin, released on Hugging Face under the Apache 2.0 license. It reads up to 524,288 tokens in one prompt, calls tools in XML or JSON, and beats Alibaba’s Qwen3.8-27B on most agentic coding tests in the lab’s own results.
The lab posted a tuned model and a separate base checkpoint over the weekend, both built for coding and multistep office work. No outside group has reproduced the benchmark scores yet. The 512K context window also has hardware costs that the spec sheet doesn’t show.
The model card calls the release the “flagship teacher model for the Nanbeige4.3 series.” The phrase describes the lab’s strategy. Nanbeige trained the large model first so it can distill coding and tool-use skills into a smaller 3B model it plans to release next.
What Is Nanbeige4.3-33B?
Nanbeige4.3-33B is a dense model, so all 33 billion parameters run on every token, with no expert routing. Many recent open releases use a mixture-of-experts design that switches on only part of the network per token, which cuts compute per request. A dense model gives up that efficiency in exchange for a simpler architecture that standard tools can fine-tune and serve.
The model handles text only, ships in BF16 precision and lists English and Chinese as its languages. Reasoning is switched on by default. For tool calls, it accepts XML or JSON, and the lab recommends XML for the best results. Deployment recipes on the model card cover Transformers, SGLang and vLLM.
The release also includes Nanbeige4.3-33B-Base, a raw checkpoint for teams that want to run their own post-training. For research groups, that second file may prove more useful than the headline scores. A base checkpoint lets a bank, a law firm or a software vendor shape the model around its own data without inheriting the lab’s chat tuning.
How Does Nanbeige4.3-33B Compare With Qwen3.8-27B?
The lab’s own figures show Nanbeige4.3-33B ahead on agentic coding by wide margins. It scored 80.9 on Terminal-Bench against 73.0 for Qwen3.8-27B, and 65.5 on DeepSWE against 42.5. Across 20 published tests against Qwen3.8-27B, Muse Glimmer 30B and Gemma 4 31B, it placed first on 17.
The test set leans toward workplace tasks. Beyond coding, Nanbeige reports results on Automation Bench, DeepPlanning and JobBench, which scores models across specific occupations.
The four models share a license but differ in size and context length.
| Model | Parameters | Context window | License |
|---|---|---|---|
| Nanbeige4.3-33B | 33B dense | 524,288 tokens | Apache 2.0 |
| Qwen3.8-27B | 27B dense plus vision encoder | 262,144 native, up to 1M | Apache 2.0 |
| Muse Glimmer 30B | about 29.6B dense | 131,072+ | Apache 2.0 |
| Gemma 4 31B | 31B | 256K | Apache 2.0 |
Figures come from each model’s own card, compiled in a side-by-side comparison.
Nanbeige’s listed maximum is about double Qwen’s native window. Qwen’s card lists extension up to 1 million tokens, though, and Qwen3.8-27B includes a vision encoder that Nanbeige4.3-33B lacks. For teams that need image input, that gap decides the choice before any benchmark does. marktechpost
Why the Benchmark Claims Need Independent Testing
Every score in Nanbeige’s table comes from Nanbeige. As of Oct. 25, no Hugging Face inference provider was listed for the tuned model, so open weights did not yet mean a hosted service was available there. Developers who want to check the claims have to download and run the model themselves.
Self-reported benchmarks reflect choices the publisher makes about tests, prompts and evaluation tools. Nanbeige notes that it evaluated Terminal-Bench 2.1 using the Terminus 2 harness. Scores can move when another harness or prompt format is used, so teams should run each candidate model through the same setup with their own tasks.
Long context raises a separate question. A 524,288-token limit tells developers how much text the model accepts. It says nothing about how well the model recalls a detail buried 300,000 tokens back. Recall at long range varies between models and should be measured at the lengths a product needs.
What Hardware Does Nanbeige4.3-33B Need?
Full-Precision Deployment
Running the full-precision Nanbeige4.3-33B is a data center job. The SGLang and vLLM examples on the model card split the model across four GPUs with tensor parallelism. A 30B-class open model, in other words, does not run at full context on a single consumer graphics card.
Quantized Builds
Compressed builds lower the bar. A community 4-bit NVFP4 version peaked at 21.64 GiB of GPU memory on an Nvidia RTX PRO 6000 Blackwell card, generating about 64 tokens per second. That test ran at a 16,384-token context, a small fraction of the maximum. Memory for the attention cache grows with context length, so a product that fills hundreds of thousands of tokens per request needs far more than a single-card budget.
The lab has not yet posted official compressed builds of the 33B model, and users on its Hugging Face page have asked for an FP8 release.
Where Nanbeige Fits Inside BOSS Zhipin
Nanbeige LLM Lab sits within Kanzhun Limited, which runs BOSS Zhipin and trades on Nasdaq and the Hong Kong exchange under the tickers BZ and 2076. Kanzhun raised $912 million in its June 2021 Nasdaq listing.
Before this release, the lab mostly built 3B-scale models. Its Nanbeige4.2-3B reused Transformer layers to add depth without adding parameters. That small model logged about 116,000 downloads on Hugging Face in its latest month. Its model card says training is underway on Nanbeige4.5 for release later in 2026.
The model card lays out the Nanbeige4.3-33B plan: build a high-capacity teacher agent first, then use it to distill advanced agent skills into smaller models. If the distilled 3B model keeps much of the larger model’s coding ability, it could run on far cheaper hardware. That is where the lab’s earlier small models found their audience.
What the Release Means for Software Teams
Open-weight releases keep pushing down what developers pay for capable language models. One analysis tracked 80% price cuts within months as open models capped what API sellers can charge. A 33B model under Apache 2.0 gives software companies another option for hosting a coding or support agent on their own servers. Apache 2.0 carries no commercial-use limits, unlike some open-weight licenses with caps tied to revenue or user counts.
Self-hosting is not always the cheaper path. A team without GPU capacity or machine learning operations staff may pay less per month for a hosted model billed per token than for a four-GPU cluster running around the clock. Companies with rules on model provenance will also need an internal review before putting any new model into production.
What to Do Next
For teams weighing Nanbeige4.3-33B, these steps cut the risk of a bad bet:
- Run it alongside your current model on 50 to 100 real tasks from your backlog, using one evaluation harness for both.
- Test recall at the context lengths your product actually uses, not at the 512K maximum.
- Use XML tool calls, as the lab recommends, before judging agent reliability.
- If you plan to fine-tune, start from the Base checkpoint and watch for the 3B release before buying multi-GPU hardware.
The lab’s numbers make Nanbeige4.3-33B worth a test run. Independent results over the coming weeks will show whether its lead over Qwen3.8-27B holds outside the lab’s own table.
Frequently Asked Questions
What is Nanbeige4.3-33B?
Nanbeige4.3-33B is a 33-billion-parameter open-weight language model from Nanbeige LLM Lab, part of Chinese recruitment company Kanzhun, which runs BOSS Zhipin. It is built for agentic coding and multistep office tasks. The lab designed it as a teacher model whose skills will be distilled into a smaller 3B model.
Is Nanbeige4.3-33B free for commercial use?
Yes. Both the tuned model and the base checkpoint are released under the Apache 2.0 license, which allows commercial use without revenue or user caps. Companies can download the weights from Hugging Face and run them on their own hardware. They still need to cover their own compute and hosting costs.
How big is the Nanbeige4.3-33B context window?
The model card lists a maximum context of 524,288 tokens, usually described as 512K. That is about double Qwen3.8-27B’s native window. A larger limit does not guarantee strong recall across the whole window, so teams should test performance at the context lengths their applications actually use.
Does Nanbeige4.3-33B beat Qwen3.8-27B?
On Nanbeige’s own benchmarks, yes. It scored 80.9 on Terminal-Bench 2.1 against 73.0 and 65.5 on DeepSWE against 42.5, placing first on 17 of 20 tests. No independent group has reproduced those results yet. Qwen3.8-27B also supports image input, which Nanbeige4.3-33B does not.
What hardware do you need to run Nanbeige4.3-33B?
Official examples for SGLang and vLLM split the full-precision model across four GPUs. A community 4-bit build used about 22 GiB of GPU memory at a 16,384-token context on a single Nvidia RTX PRO 6000 card. Longer contexts need much more memory for the attention cache.
Who makes Nanbeige models?
Nanbeige LLM Lab belongs to Kanzhun Limited, operator of the BOSS Zhipin recruitment platform. Kanzhun trades on Nasdaq as BZ and in Hong Kong as 2076. Before this release, the lab focused on small 3B-scale models, including Nanbeige4.2-3B, and says Nanbeige4.5 is in training.
