Nanbeige releases dense 33B coding model under Apache 2.0

What was released

Chinese lab Nanbeige has released the weights of Nanbeige4.3-33B on Hugging Face. The model has 33B parameters and is dense rather than a mixture-of-experts (MoE). It is licensed under Apache 2.0, which allows commercial use. The upload appears in the repository's commit history.

Most of the large open-weight models released over the past month have been MoEs. Meta's Muse Spark Open has 816B total and 42B active parameters. Reflection AI's Beam has 501B total and 23B active, and MiniMax's M3.1 Flash has about 428B total and 23B active. In each of these, only part of the model runs for each token. Nanbeige went the other way with a dense design, in which all 33B parameters run every time.

Self-reported score: 80.9 on Terminal-Bench 2.1

The model card reports a score of 80.9 on Terminal-Bench 2.1. Nanbeige says the model beats open models of similar size on most benchmarks.

Earlier reported scores on the same benchmark help put that number in context. All of them are self-reported by the companies.

• Reflection AI's Beam (501B total, 23B active): 80.1% (reported by VentureBeat; secondary source).
• Meta's Muse Spark Open (816B total, 42B active): 76.4 (Meta's model card).
• Recursive's Flint (40B total, about 2.5B active): 54.9% (company figure; secondary source).

Taken at face value, the scores put a dense 33B model level with MoEs that have 15 to 25 times as many total parameters. But companies do not necessarily measure the same way: harnesses, number of attempts and time limits can all differ. Lining up self-reported numbers from different companies does not establish a ranking.

No third-party evaluation yet

As of this writing, third parties such as Artificial Analysis, Arena and Epoch have not yet scored the model. Company figures and third-party figures often differ. Anthropic reported 39.2% for Haiku 5.5 on Terminal-Bench 4.0, but Artificial Analysis measured 33%. Beam looked strong on its own numbers but scored just 3.0% on Artificial Analysis's Terminal-Bench 4.0. It is not yet known whether the 80.9 on Terminal-Bench 2.1 will hold up under other methods or on newer versions of the benchmark.

What to watch

For a dense 33B model, the compute used per request is about the same as the active parameters of the MoEs listed above (23B to 42B). The full set of weights, though, is far smaller than an MoE's. Quantized, it can fit on fewer accelerators, which suits running it locally or fine-tuning it. If the company's numbers hold up, developers working on harness engineering get another option that is easy to work with.

There are three things to watch. First, whether Artificial Analysis publishes an Intelligence Index score and its own Terminal-Bench results. Second, how the model compares with leading open-weight models such as DeepSeek V4.1-Flash-1017 (45) and Kimi K3.1 Pro (55), whose weights are due by November 4. Third, how much safety evaluation Nanbeige discloses. On October 16, Epoch AI estimated that the cyber capabilities of open-weight models are six to seven months behind closed models. A strong coding-agent model released under Apache 2.0 is just as available to attackers as to defenders. For now, the only numbers available are Nanbeige's own.