What it does
Nanbeige4.3-33B is an open-weight 33B language model from the Nanbeige team, built for agentic coding and professional "cowork" tasks such as office automation. It is the flagship and teacher model of the Nanbeige4.3 series, meant to later distil its agent skills into smaller models such as a planned Nanbeige4.3-3B. It is a dense Llama-architecture text-generation model (52 layers, BF16 weights) with a 512K-token context, released with a separate Base model under Apache 2.0.
Official source 1 ↗ · Official source 2 ↗Features
- Agentic coding scores (self-reported). In Nanbeige's own comparison with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, it scores 80.9 on Terminal-Bench 2.1 (Terminus 2 harness), 65.5 on DeepSWE, 60.3 on SWE-Bench Pro and 53.0 on NL2Repo, which was run inside Claude Code. These are vendor-reported numbers, not independent tests. Official source 1 ↗
- Office and professional "cowork" agents. Nanbeige also targets multi-step work tasks. It reports 54.2 on Automation Bench, 44.8 on JobBench (ahead of the baseline in most of the 35 occupations tested) and 84.8 on GDPval, all self-reported. Official source 1 ↗
- 512K-token context. The model accepts up to 524,288 tokens of context, and the card recommends up to 65,536 new tokens for agent and tool-use tasks and 262,144 for reasoning and chat. Official source 1 ↗
- Thinking on or off, kept across turns. Reasoning is on by default and can be switched off per response with enable_thinking. preserve_thinking keeps earlier reasoning in multi-turn conversations; Nanbeige recommends turning it on for tool use, office tasks and code-agent workflows. Official source 1 ↗
- Tool calling in XML or JSON. Passing tools switches the chat template to tool use. XML tool calls are recommended for best results and JSON is supported; the vLLM example enables automatic tool choice. Official source 1 ↗
- Reasoning benchmarks (self-reported). Nanbeige reports 95.2 on AIME 2026, 89.5 on GPQA Diamond and 84.0 on the AA-LCR long-context reasoning test. Official source 1 ↗
- Base model for your own post-training. Nanbeige4.3-33B-Base is released separately as a starting point for training general or domain-specific agents. Nanbeige reports it beating Gemma4-31B-Base on all eight listed base-model benchmarks, for example 66.9 vs 59.3 on MMLU-Pro and 74.6 vs 72.5 on MBPP. Official source 1 ↗
Pricing & free limits
The post-trained agentic model with thinking and tool-calling chat template, as BF16 safetensors on Hugging Face. You pay only for the GPUs that run it. Official source 1 ↗ · Official source 2 ↗
The pre-trained base model without the agentic post-training, intended for your own fine-tuning or post-training. Official source 1 ↗
Billing details
Billing details were not established by the checked official sources.
Free access & limits
Free access and its limits were not established by the checked official sources.
Best for & limitations
Self-hosting the model behind a coding agent. Nanbeige ran its coding benchmarks through agent harnesses such as Terminus 2, mini-swe-agent and Claude Code, and recommends keeping reasoning across turns for code-agent workflows. Official source 1 ↗
Automating multi-step professional and office tasks with tools, the "cowork" workload Nanbeige measures across many occupations. Official source 1 ↗
Training your own general or domain-specific agent from the Base model; Nanbeige also invites collaborations on industry-focused agents. Official source 1 ↗ · Official source 2 ↗
- Benchmarks are run by Nanbeige itself. All scores come from Nanbeige's own runs, and no independent evaluations are cited on the card. Nanbeige used the same settings and harness for every model. Some use an AI judge or revised rubrics, such as SWE-Atlas-QnA (judged by GPT-6 Sol) and GDPval. Official source 1 ↗
- Not ahead on every test. In Nanbeige's own table, Qwen3.8-27B scores higher on SWE-Bench Pro (61.7 vs 60.3) and GDPval (86.4 vs 84.8), and Gemma4-31B scores higher on GPQA Diamond (91.3 vs 89.5). Official source 1 ↗
- Needs multi-GPU hardware. The official weights are BF16, and the card's SGLang and vLLM examples split the model across 4 GPUs. The card gives no single-GPU or quantized setup. Official source 1 ↗ · Official source 2 ↗
- Custom code and limited languages. The Transformers example requires trust_remote_code=True, and the card lists only English and Chinese. Official source 1 ↗
- Smaller distilled models are not out yet. Nanbeige presents this as the teacher for a coming distillation pipeline; small models such as Nanbeige4.3-3B are announced as upcoming, not released. Official source 1 ↗
Submit a tool
