Nanbeige4.3-33B: a 33-billion-parameter open model for coding agents, under the Apache-2.0 license
Released on Hugging Face with its weights and a 512,000-token context, the model beats similarly sized rivals on most of the tests published by its developer, but not all of them.
The Nanbeige lab has published Nanbeige4.3-33B on Hugging Face, a 33-billion-parameter language model whose weights are distributed under the Apache-2.0 license, one of the most permissive: commercial use, modification and redistribution are allowed. The base version, before post-training, is also available. According to the repository history, it was created on October 24 and its model card updated on October 25.
A “teacher” model for agents
The developer presents this model as the teacher model of its 4.3 series: it is meant to distill agentic capabilities into smaller models, starting with a 3-billion-parameter version announced without a date. It targets two uses: agentic coding (a model that runs commands, reads the results and fixes things) and automated office tasks.
On the technical side, the model card lists a context window of 524,288 tokens, a switchable reasoning mode and a tool-calling format (XML recommended). The model runs with Transformers, SGLang or vLLM; the BF16 weights weigh about 66 GB, and the deployment examples provided spread the model across four GPUs.
Scores worth reading in full
Nanbeige compares its model with three open competitors of similar size: Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all evaluated, according to the developer, with the same settings and the same harness. It claims 80.9 on Terminal-Bench 2.1 (versus 73.0 for Qwen3.8-27B), 23.8 on Terminal-Bench 4.0 (6.3 for Qwen) and 65.5 on DeepSWE (42.5 for Qwen).
The published table also shows its limits: on SWE-Bench Pro, Nanbeige4.3-33B scores 60.3, behind Qwen3.8-27B (61.7); on GDPval, 84.8 versus 86.4 for Qwen; on GPQA Diamond, 89.5 versus 91.3 for Gemma4-31B. All these figures come from the developer; no independent measurement has been published yet.
Who is it for?
A 33-billion-parameter model under Apache-2.0 appeals to teams that want a self-hosted coding agent without depending on an API. The comparisons should still be read with caution: agent scores vary widely depending on the tooling used (the “harness”), and Nanbeige’s are not directly comparable with figures published by other developers using other tools.
Our take
The license and the size are the real strengths: a truly open model in this class remains rare for coding agents. The reported gaps on Terminal-Bench and DeepSWE are large, but they come from a single developer. The fact that the model card also shows tests where the model is beaten speaks to its seriousness; independent evaluations are still needed.
Sources
- Hugging Face — Nanbeige/Nanbeige4.3-33B model card Primary source
- Hugging Face — Nanbeige4.3-33B repository API (creation and update dates) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
