Nanbeige opens 33B weights for self-hosted coding and office agents

Nanbeige has released the weights for Nanbeige4.3-33B, a 33-billion-parameter model aimed at coding and office-work agents. Its official model card, updated October 25, presents comparisons with other open models of similar size. Developers can host the model themselves as the engine for agents that work with code and business documents.

The release includes both a model adapted for conversations and tool use and a separate base checkpoint for further training. That gives organizations a choice between starting with an agent-oriented model and adapting a foundation to their own tasks. The base model’s documentation explicitly identifies general-purpose and domain-specific agents as post-training uses.

The published conversation template is text-oriented. For image, video or audio input, it inserts a reminder that the model cannot process that medium directly. An agent working with spreadsheets or scanned documents therefore needs tools that expose their contents as text. The template supports requesting named tools with arguments and feeding their results back into the conversation.

In Nanbeige’s tests, the model scored 80.9 on Terminal-Bench 2.1, compared with 73.0 for Qwen3.8-27B. On JobBench, the respective scores were 44.8 and 34.5. The lab says it used each model’s highest thinking effort and matched inference settings and evaluation tools within each benchmark. These are results reported by the model’s developer.

JobBench tests work that requires producing a deliverable from multiple sources. Its currently published main dataset contains 65 tasks spanning 35 occupations. Each task comes with a request, reference files and a weighted grading rubric; some require finding additional material through search. The source material can mix tables with explanatory documents, requiring an agent to reconcile different file types.

The researchers behind that benchmark explained in their May paper that task selection starts with work professionals want to delegate. Their scoring method awards a rubric’s weight when every linked criterion passes. Finding the right number while making an incorrect intermediate judgment can lose the entire chain’s credit. One reporter task in the paper combines water-quality tables from multiple years with public guidance to identify threshold exceedances and prepare an editorial package. The new model’s score describes performance within such defined materials and grading criteria.

Server resources for long context and further training

The released agent model’s configuration allows up to 524,288 tokens of context. Hugging Face lists precisely 33,118,977,024 parameters stored in BF16. At two bytes per parameter, the weights alone represent approximately 66.2 billion bytes, or 66.2 GB in decimal units. Conversation storage and intermediate computation require additional memory when running the model.

NVIDIA’s NeMo AutoModel documentation describes a further-training route. The current Nightly documentation supports loading the model through its native implementation and supplies two recipes. Full-parameter training uses two servers with eight H100 GPUs each. A LoRA recipe uses one eight-GPU server.

NVIDIA reports runs of 100 steps on H100 80 GB GPUs, with step times of roughly 27 seconds for full training and 36 seconds for LoRA. These records describe the specified training configurations. The accompanying loss values measure training behavior rather than held-out task quality.

The base checkpoint also has a different context setting: NVIDIA documents 131,072 maximum positions. Its separate card compares eight knowledge, mathematics and coding benchmarks with Gemma4-31B-Base. Nanbeige reports GPQA scores of 62.0 and 41.5 respectively. That comparison concerns the foundations available for subsequent adaptation.

Both model cards identify Apache 2.0 as the license. Its terms permit reproduction, modification and distribution subject to conditions, including providing the license, retaining applicable notices and identifying changed files when redistributing. Those obligations accompany redistributed adaptations. The license also requires preservation of applicable copyright, patent and attribution notices. When an original distribution includes a NOTICE file, redistributed derivatives must carry its relevant attribution notices in a readable form.