Nanbeige released Nanbeige4.3-33B, an open model for agentic coding

Nanbeige LLM Lab, the AI team inside Chinese recruitment platform BOSS Zhipin, has released Nanbeige4.3-33B on Hugging Face under the Apache 2.0 license. It is a 33 billion parameter dense model built for agentic coding and multi-step office tasks, with a 524,288 token (512K) context window. The lab also published a Base version for teams that want to do their own post-training.

The headline claim comes from the lab's own model card. MarkTechPost counts it leading 17 of 20 benchmark rows against three open models near its size: Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. Reported highlights include 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen3.8-27B, and 65.5 on DeepSWE against 42.5. It also posts 54.2 on Automation Bench versus 38.3. It trails Qwen3.8-27B on SWE-Bench Pro, 60.3 to 61.7.

Treat those numbers with care. All are self-reported and unverified. MarkTechPost found several competitor figures in Nanbeige's table that differ from what those labs publish, and an analysis by OrcaRouter notes that four of the wins against Qwen are by less than one point. The large gaps are on long, multi-step agent tasks, which is where the model appears genuinely strong.

Under the hood it is a standard Llama style architecture with no mixture of experts, so all 33B parameters run on every token. The chat template supports a thinking toggle, a setting to keep earlier reasoning across turns for tool use, and tool calling. MarkTechPost estimates the BF16 weights need about 66 GB, and the card's server recipes use four GPU tensor parallelism with SGLang or vLLM. Startup Fortune notes a 4-bit quantization could fit on a single 24 GB card, with limited context. Nanbeige calls it a teacher model and plans to distill it into smaller models, starting with a 3B version.

For developers, this is a permissively licensed coding agent model you can host yourself with no commercial limits. That matters as GPU rental prices rise and teams weigh API bills against owning their stack.

What to try: run it on your own agent tasks rather than trusting the table. Compare it with Qwen3.8-27B on real repository work, watch memory use at long contexts, and wait for independent benchmark runs before switching production workloads.

Building with AI? See this week's hot AI repos and how to use each one.