Nanbeige Releases 33B Open Agentic Coding Model Nanbeige4.3-33B
Nanbeige LLM Lab released Nanbeige4.3-33B, a 33B open dense model for agentic coding and office cowork tasks.
The model uses BF16 weights and runs on multi-GPU servers; its SGLang and vLLM recipes use 4-way tensor parallelism. MarkTechPost calculated that BF16 weights need about 66 GB, and the model page lists one community quantization. Nanbeige positions Nanbeige4.3-33B as the flagship “teacher” model of the Nanbeige4.3 series. The lab’s stated strategy is to build a strong teacher first and then distill agentic skills into smaller models, starting with Nanbeige4.3-3B.
The config.json declares a standard LlamaForCausalLM architecture, with no expert routing, so all 33B parameters run on every token. It has 52 layers, hidden size 7,168, 56 query heads and 8 key-value heads with grouped-query attention, head dimension 128, FFN intermediate size 22,016, a vocabulary of 166,912 tokens, RoPE theta of 120,000,000 and 524,288 max positions. The chat template exposes three controls. enable_thinking toggles reasoning and is on by default; preserve_thinking keeps reasoning from earlier turns and Nanbeige recommends it for multi-turn tool use and code agents; passing tools enables tool calling, with an XML call format recommended.
For its benchmark table, Nanbeige compared the model with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B. Every model ran at its highest thinking effort, with shared hyperparameters and harnesses, and all scores in this section are self-reported. Nanbeige’s table shows the model leading 17 of 20 rows. Its best result is 80.9 on Terminal-Bench 2.1, against 73.0 for Qwen3.8-27B. Its lowest score is 23.8 on Terminal-Bench 4.0. It also trails Qwen3.8-27B on SWE-Bench Pro, 60.3 versus 61.7.
On DeepSWE, Nanbeige4.3-33B scores 65.5 versus 42.5 for Qwen3.8-27B. On Terminal-Bench 4.0, it scores 23.8 while Gemma4-31B scores 0.0. In agentic cowork tests, it posts 54.2 on Automation Bench versus 38.3 and 44.8 on JobBench versus 34.5; on YC-bench, simulated earnings reach $1,025,437 versus $641,290. In reasoning, it reports 95.2 on AIME 2026 and 84.0 on AA-LCR.
It loses three of the 20 rows in the lab’s table. Qwen3.8-27B leads SWE-Bench Pro at 61.7 versus 60.3 and GDPval at 86.4 versus 84.8, while Gemma4-31B leads GPQA Diamond at 91.3 versus 89.5. The Base model beats Gemma4-31B-Base on all eight reported tasks, including MMLU-Pro at 66.9 versus 59.3 and GPQA at 62.0 versus 41.5.