Nanbeige Releases Nanbeige4.3-33B, a 33B Open Agentic Coding Model
MarkTechPost Asif Razzaq
Nanbeige just open-sourced a 33B coding model with a 512K context window. It’s meant to act like a teacher model first, then pass its skills down to smaller ones.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nanbeige LLM Lab has put Nanbeige4.3-33B on Hugging Face under Apache 2.0, and it’s not a tiny demo model pretending to be useful. This is a 33B dense model built for agentic coding and office-style cowork tasks, with a 524,288-token context window and BF16 weights. The lab also released Nanbeige4.3-33B-Base for post-training work.
The model comes out of the BOSS Zhipin recruitment platform’s lab, which makes the whole release feel less like a research flex and more like a product-minded push. Nanbeige says this is the flagship teacher in the Nanbeige4.3 series. The plan is straightforward: make the strong model first, then distill its agentic skills into smaller versions, starting with Nanbeige4.3-3B.
On its own table, Nanbeige says the model leads 17 of 20 benchmarks against three open models near the same size: Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. Its strongest cited result is 80.9 on Terminal-Bench 2.1, ahead of Qwen3.8-27B’s 73.0. DeepSWE shows a wider gap, 65.5 versus 42.5. But the model does not win everywhere. It trails Qwen3.8-27B on SWE-Bench Pro, 60.3 to 61.7.
The release is also unusually explicit about how the model works. There is no MoE routing; all 33B parameters run on every token. The chat template exposes controls for thinking and tool use, and Nanbeige recommends an XML call format for tools. Deployment guidance points to Transformers, SGLang and vLLM, with 4-way tensor parallelism. The lab also says no Hugging Face Inference Provider is hosting it yet.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of open release: a serious model, a clear license, and enough deployment detail to make it actually usable. The only part that should make anyone squint is the benchmark table, because once one lab starts publishing its own head-to-head scorecard, everyone else gets to bring a calculator and a raised eyebrow.
Read more about this at: MarkTechPost
