Nanbeige has released Nanbeige4.3-33B, a 33-billion-parameter model aimed at coding agents and multi-step office work. Its public repository was created on October 24, with downloadable weights available under Apache 2.0.
A self-hosting option for agent workflows
The model supports a 512K-token context window, configurable reasoning and tool calling. Deployment examples cover Transformers, vLLM and SGLang, with server examples distributing the model across four GPUs.
That is an example configuration, not a universal minimum hardware requirement. This is a self-hosting option, not a lightweight phone model; teams still need to size their infrastructure for the intended workload.
The benchmark picture is mixed
Nanbeige reports 80.9 on Terminal-Bench 2.1, versus 73.0 for Qwen3.8-27B. But Qwen leads on SWE-Bench Pro in the same comparison.
These are vendor-run results, not independent proof of better performance in your repository. A strong result on one coding benchmark is a reason to evaluate the model, not a reason to assume it will lead on every maintenance task.
Why it matters for builders
For developers who want control over their agent infrastructure, this adds another openly licensed model to evaluate on real workflows. Deployment control also means owning the serving setup, resource budget and operational checks.
For a small app team such as SunMarc, a useful trial would be a bounded bug-fix task: compare successful fixes, inference cost and human review time against the current setup. Use the same task and acceptance checks so that a cheaper run does not hide a less useful result.
Our coverage of Recursive Flint for local coding agents examines another deployment choice. The useful comparison is the complete workload and its hardware requirements—not just whether weights can be downloaded.