Nanbeige LLM Lab, the AI research group inside Chinese recruiting platform BOSS Zhipin, has open-sourced Nanbeige4.3-33B, a 33-billion-parameter model built for agentic coding and professional "cowork" tasks. The weights are on Hugging Face under an Apache 2.0 license, and the lab's own benchmark table puts the model ahead of Qwen3.8-27B on 18 of the 20 tests it reports.
Nanbeige4.3-33B ships with a 524,288-token context window, roughly 512K, and reasoning switched on by default. The model card lists it as a fine-tune of Nanbeige4.3-33B-Base, published in BF16 weights.
The lab's headline claim is "state-of-the-art agentic performance among open models in the ~30B parameter class." Its comparison set is Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. Nanbeige says every model was run at its highest thinking effort, with the same harness and inference settings, and that its own results used a preserve-thinking setting.
Key facts
- 01Nanbeige LLM Lab released Nanbeige4.3-33B on Hugging Face under an Apache 2.0 license.
- 02The model leads Qwen3.8-27B on 18 of 20 self-reported benchmarks, including 80.9 vs. 73.0 on Terminal-Bench 2.1.
- 03Its context window is 524,288 tokens, with reasoning enabled by default.
- 04Qwen3.8-27B leads on SWE-Bench Pro (61.7 vs. 60.3) and GDPval; Gemma4-31B leads on GPQA Diamond (91.3 vs. 89.5).
- 05The 33B model is the series' teacher, meant to be distilled into smaller models such as a planned Nanbeige4.3-3B.
The widest margins are on terminal and software-agent benchmarks. Nanbeige4.3-33B scores 80.9 on Terminal-Bench 2.1, against 73.0 for Qwen3.8-27B, 61.8 for Muse-Glimmer-30B and 47.2 for Gemma4-31B. On the harder Terminal-Bench 4 it posts 23.8, versus 6.3 for Qwen3.8-27B and 0.0 for Gemma4-31B. On DeepSWE it scores 65.5 to Qwen's 42.5.
The workflow-style tests follow the same pattern. Nanbeige reports 54.2 on Automation Bench (Qwen3.8-27B: 38.3), 44.8 on JobBench (34.5), 63.3 on DeepPlanning (49.2) and 70.4 on Toolathlon-Verified (68.5). On YC-bench, which scores in simulated dollars, Nanbeige4.3-33B finishes with $1,025,437 against $641,290 for Qwen3.8-27B. The card adds that on JobBench the model beats the baseline in most of 35 occupations.
Reasoning scores are closer. Nanbeige4.3-33B posts 95.2 on AIME 2026, against 94.5 for Qwen3.8-27B, 94.7 for Muse-Glimmer-30B and 89.2 for Gemma4-31B. It also reports 52.6 on SciCode and 84.0 on the AA-LCR long-context test, both the top marks in the table.
The table is not a clean sweep. Qwen3.8-27B leads on SWE-Bench Pro, 61.7 to 60.3, and on GDPval, 86.4 to 84.8. Gemma4-31B tops GPQA Diamond at 91.3 to Nanbeige's 89.5. All of the figures are self-reported, and the lab has not cited any independent reproduction.
For deployment, Nanbeige4.3-33B runs on Transformers, SGLang or vLLM. Its chat template supports enable_thinking and preserve_thinking flags, the latter keeping reasoning across turns, which the lab recommends for tool use and agent work. Tool calls can be written in XML, the recommended format, or JSON.
The model card recommends a temperature of 1.0 and up to 65,536 new tokens for agentic and tool-use tasks, and a temperature of 0.6 with up to 262,144 new tokens for reasoning and chat. The card does not disclose training data sources, training method or compute.
Nanbeige describes the 33B release as the "flagship teacher model" for the Nanbeige4.3 series. According to the model card, its job is to anchor distillation into smaller models, starting with a planned Nanbeige4.3-3B.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
