Nanbeige Releases Nanbeige4.3-33B, a 33B Open Agentic Coding Model

Nanbeige LLM Lab, the AI group inside the recruitment platform BOSS Zhipin, has released Nanbeige4.3-33B on Hugging Face under Apache 2.0. It's a 33B dense model aimed at agentic coding and office-style cowork tasks, with a 524,288-token context window. The lab also shipped a Base model for teams that want to do their own post-training.

In the lab's own comparison against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, it leads 17 of 20 benchmarks, including 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen. If those numbers hold, it's one of the strongest open coding agents in its size class.

We like this release on its fundamentals. A permissive license, open weights, a Base checkpoint, a long context, and clear deployment recipes for Transformers, SGLang and vLLM. That's the kind of package that lets companies run a capable code agent on their own hardware instead of sending their codebase to someone else's API. The plan to use this as a teacher and distill agentic skills into smaller models, starting with a 3B version, is also the right direction. Small, capable agents are where a lot of practical value will land.

Now the caution. Every score here is self-reported, and MarkTechPost found several competitor figures in Nanbeige's table that don't match what those labs publish. Google lists Gemma 4 31B at 84.3 on GPQA Diamond; Nanbeige shows 91.3. The SWE-Bench Pro numbers for Muse-Glimmer and Gemma look swapped. Different harnesses can explain some gaps, but not all of them. Until independent runs land, we'd treat the 17-of-20 claim as a marketing number, not a ranking.

There's also an honesty point in its favor: the table shows where it loses, including SWE-Bench Pro, where Qwen3.8-27B edges it 61.7 to 60.3, and a 23.8 on Terminal-Bench 4.0. That harder benchmark is a useful reminder that even strong open agents still fail most of the toughest terminal tasks.

On cost, this isn't a laptop model. It's dense, so all 33B parameters run on every token, and by MarkTechPost's arithmetic the BF16 weights need about 66 GB, with the full context adding roughly 112 GB of KV cache. The card's server recipes assume four GPUs. Accessible for a company, not for most hobbyists.

Where this leads if it scales: capable autonomous coding agents become a commodity anyone can self-host. That's good for competition and privacy. It also means agents with terminal access spread faster than the guardrails around them. I'm watching whether teams pair these models with real sandboxing, not just good benchmark scores.

Agentic coding benchmarks like Terminal-Bench and DeepSWE test whether a model can operate tools and complete multi-step software tasks, not just write snippets. The model's chat template exposes thinking on by default, an option to preserve reasoning across turns for tool use, and tool calling with a recommended XML format. At publication, no Hugging Face Inference Provider was hosting it.