Nanbeige Opens 33B Model Under Apache 2.0 With 512K Context

Nanbeige LLM Lab has posted Nanbeige4.3-33B on Hugging Face under Apache 2.0, a 33B-parameter model with a 524,288-token context window and reasoning enabled by default. The lab calls it the "flagship teacher model for the Nanbeige4.3 series" and claims "state-of-the-art agentic performance among open models in the ~30B parameter class."

The model card's benchmark table puts Nanbeige4.3-33B ahead of Qwen3.8-27B on most agentic coding tasks: 80.9 versus 73.0 on Terminal-Bench 2.1, 65.5 versus 42.5 on DeepSWE, 52.6 versus 46.7 on Scicode. Qwen3.8-27B edges it on SWE-Bench Pro, 61.7 to 60.3. On reasoning, Nanbeige reports 95.2 on AIME 2026 against Qwen's 94.5, though Gemma4-31B leads GPQA Diamond at 91.3 to 89.5. All of these figures are self-reported, with no independent reproduction attached to the release.

The deployment settings in the card point at serious tool use: SGLang and vLLM commands with tensor-parallel size 4, a Qwen3-coder tool parser, and a 262,144-token max for reasoning tasks. A "preserve_thinking" flag retains reasoning across turns, which the lab recommends keeping on for tool-use and agent tasks.

Model-card benchmark claims from Chinese open-source releases have been a running thread in our coverage, with Moonshot's Kimi K3.1 Pro landing today and IQuest-Q1 targeting the same Terminal-Bench 2.1 earlier in the week.