Nanbeige releases open 33B model for agentic coding and cowork

Nanbeige LLM Lab released Nanbeige4.3-33B on Hugging Face under the Apache 2.0 license.[1] The lab describes the model as a dense language model with 33 billion parameters, built for agentic coding and office-style cowork tasks.[1] Nanbeige calls it the flagship teacher model for the Nanbeige4.3 series.[2] The release also includes a Base version intended for custom post-training, giving teams a separate option if they want to adapt the model rather than use the flagship version as released.[1]

The model supports a context window of 524,288 tokens.[1] Nanbeige says it plans to use the model as a teacher for smaller models, beginning with Nanbeige4.3-3B.[3] The lab also says Nanbeige4.3-33B can tackle real-world tasks when integrated with mainstream agent harnesses, and describes inference as cost-efficient.[3] Those statements outline the lab’s intended uses, rather than establishing how the model will perform in a particular team’s tools or workload; professionals evaluating it should test their own tasks and setup.

In its own comparison table, Nanbeige reports that the model leads 17 of 20 benchmarks against three open models near the 30-billion-parameter range.[1] The lab reports a score of 80.9 on Terminal-Bench 2.1, compared with 73.0 for Qwen3.8-27B.[1] Its table also reports scores of 65.5 on DeepSWE, 54.2 on Automation Bench and 44.8 on JobBench.[1] The comparison is not a clean sweep: on SWE-Bench Pro, Nanbeige’s score is 60.3, below Qwen3.8-27B’s 61.7.[1]

Nanbeige says it evaluated all models at their highest thinking effort and used the same inference hyperparameters and harness for each benchmark.[2] The scores in the reported section are self-reported, and some competitor figures in Nanbeige’s table differ from numbers published by those labs.[1] MarkTechPost reports that Nanbeige LLM Lab is part of recruitment platform BOSS Zhipin.[1] For a purchasing or deployment decision, treat the published results as a starting point: reproduce relevant tests where possible and compare performance on the tasks, constraints and tools that matter to your team.