Nanbeige lab releases open model Nanbeige4.3-33B for coding agents: 80.9 on Terminal-Bench 2.1 vs 73.0 for Qwen3.8-27B, by its own measurements
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
A mid-sized open model is catching up with closed flagships in coding by its authors' account; independent measurements will show whether that holds.
According to the model card on Hugging Face, Nanbeige4.3-33B is a dense 33-billion-parameter model under the Apache 2.0 license with a 524,288-token context, for English and Chinese; it is described as the "flagship teacher model" of the series, on which the lab plans to train smaller models, for example Nanbeige4.3-3B. The developer compared it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at maximum thinking effort and says it leads on most tests: Terminal-Bench 2.1 — 80.9 vs 73.0 for Qwen3.8-27B, DeepSWE — 65.5 vs 42.5. On some tests it trails: SWE-Bench Pro — 60.3 vs 61.7 for Qwen3.8-27B, GPQA Diamond — 89.5 vs 91.3 for Gemma4-31B. All figures come from the developer. A short account of the release is also on MarkTechPost (per the feed description, text paywalled).
Sources 2