Nanbeige releases Nanbeige4.3-33B, an open-weights model aimed at coding and office agents
The Nanbeige team published Nanbeige4.3-33B on Hugging Face under the Apache 2.0 license, a roughly 33-billion-parameter model with a 512K-token context window that it says leads open models of about 30B parameters on most agentic coding and work-task benchmarks.
What happened
- On October 24, 2026, Nanbeige published the Nanbeige4.3-33B model and its base model on Hugging Face under the Apache 2.0 license, with English and Chinese listed as languages.
- The model card describes it as the flagship teacher model of the Nanbeige4.3 series, built to later distill agent skills into smaller models such as a planned Nanbeige4.3-3B.
- Nanbeige's own comparison against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B shows it ahead on most agentic coding and work-task benchmarks, for example 80.9 on Terminal-Bench 2.1, while Qwen3.8-27B scores slightly higher on SWE-Bench Pro and GDPval.
- It supports up to 524,288 tokens (512K) of context, has optional thinking and tool-calling modes, and runs with Transformers, SGLang or vLLM.
Why it matters: A permissively licensed, mid-sized open model gives businesses and developers another self-hosted option for coding and agent work, though the benchmark results are the developer's own and haven't been independently checked here.
Sources
- Nanbeige4.3-33B model card (Hugging Face) huggingface.co
Checked October 26, 2026 against the sources above. Claims about performance or benchmarks are the company's own unless stated otherwise.