Nanbeige4.3-33B is a new open agent model under Apache-2.0

Nanbeige published Nanbeige4.3-33B, an open-weight language model with about 33 billion parameters, on Hugging Face on October 24. The weights are under the Apache-2.0 license, so companies can use them commercially. The lab says it is the best open model of its size for agentic coding and office-style tasks, and its own benchmark table mostly backs that up.

Models in this size range are the ones developers can run on a single high-end machine. If the numbers hold up outside the lab's own tests, that makes the model a candidate for self-hosted coding agents.

What the model card says

Nanbeige calls the model the "flagship teacher" of its 4.3 series. The plan is to use it to train smaller models, and the card names a Nanbeige4.3-3B as coming later. The model supports a context window of up to 524,288 tokens, which is 512K, and handles English and Chinese.

The lab compares it with three recent models of similar size: Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. All were run at their highest reasoning setting, with the same settings and test harness for each, according to the card.

Benchmarks: ahead on most, behind on a few

On the lab's table, Nanbeige4.3-33B comes first on most agentic coding tests:

• Terminal-Bench 2.1, which tests agents working in a command line: 80.9, against 73.0 for Qwen3.8-27B, the next best.
• DeepSWE: 65.5, against 42.5 for Qwen3.8-27B.
• Terminal-Bench 4.0: 23.8, against 6.3.
• AIME 2026, a math test: 95.2, against 94.7 for Muse-Glimmer-30B.

It does not win everywhere. Qwen3.8-27B scores higher on SWE-Bench Pro (61.7 against 60.3) and on GDPval, a test of professional tasks (86.4 against 84.8). Gemma4-31B leads on GPQA Diamond, a graduate-level science test, with 91.3 against 89.5.

These are the lab's own results, and the footnotes show they depend on the setup. Different tests use different agent harnesses, one grading step uses GPT-6 Sol as the judge, and the Nanbeige score uses a chat-template option that keeps the model's earlier reasoning in multi-turn use. No independent evaluation was available when this was written.

How to run it

The card shows how to load the model with Transformers, SGLang or vLLM. For agentic work it recommends a temperature of 1.0 and top_p of 0.95, with up to 65,536 new tokens. For reasoning and chat it recommends a temperature of 0.6. It also recommends XML-format tool calls for the best tool-use results.

What we don't know yet

The card links no technical report, and it gives no training data details or compute figures. The model's hosted page showed zero downloads shortly after upload, so community reports on speed and quality have not started. The 3B model that the lab promises has no date.

Watch for independent runs on Terminal-Bench and SWE-bench in the coming days. They will show whether the lead survives a different harness.