Nanbeige Open-Sources a 33B Agentic Coding Model, Claiming Benchmark Wins Over Qwen3.8

What happened

The model card defines Nanbeige4.3-33B as the series' flagship teacher model: build a high-capacity teacher agent first, then distill agentic capabilities into smaller models, including the planned Nanbeige4.3-3B. MarkTechPost reported on Oct 25 that the lab also released Nanbeige4.3-33B-Base for custom post-training.

The deployment settings point to heavy tool use: reasoning is enabled by default, and both the SGLang and vLLM recipes use 4-way tensor parallelism; the card recommends temperature 1.0, top_p 0.95 and 65,536 max new tokens for agentic and tool tasks, and 262,144 max new tokens for reasoning and chat tasks. The card's contact address uses the kanzhun.com domain, and multiple reports place the lab inside BOSS Zhipin.

The benchmark comparison sits in section 2 of the model card: all models were evaluated at their highest thinking effort under the same harness, against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. Against Qwen3.8-27B, the card's table leads on 18 of 20 items, trailing only on SWE-Bench Pro and GDPval; MarkTechPost and AI Weekly both note that all figures are lab-reported, with no independent reproduction yet.

Key facts

1. Model and license: Nanbeige4.3-33B, a 33-billion-parameter dense model under the Apache 2.0 license, with a 524,288-token context window and reasoning enabled by default.
2. Positioning: Flagship teacher model of the Nanbeige4.3 series, to be distilled into smaller models including the planned Nanbeige4.3-3B; a Base version was released alongside for custom post-training.
3. Deployment: Both SGLang and vLLM recipes use 4-way tensor parallelism; BF16 weights run about 66 GB (press arithmetic), with a community quantization listed.
4. Coding benchmarks: Terminal-Bench 2.1 at 80.9 (Qwen3.8-27B at 73.0) and DeepSWE at 65.5 (against 42.5), all lab-reported.
5. Where it trails: Against Qwen3.8-27B it trails on two items: SWE-Bench Pro, 60.3 to 61.7, and GDPval, 84.8 to 86.4.
6. Reasoning benchmarks: GPQA Diamond at 89.5 (Gemma4-31B at 91.3) and AIME 2026 at 95.2 (Qwen3.8-27B at 94.5), all lab-reported.
7. Nature of the figures: All models were evaluated at their highest thinking effort under the same harness; there is no independent reproduction yet and the card attaches no third-party verification.