Nanbeige LLM Lab (inside recruitment platform BOSS Zhipin) published Nanbeige4.3-33B on Hugging Face as a 33B dense open model under Apache 2.0, with a matching Base checkpoint for post-training. The lab positions it as the flagship teacher of the Nanbeige4.3 series for agentic coding and office-style cowork, and says it plans to distill those skills into smaller models such as Nanbeige4.3-3B. Context length on the card is up to 524,288 tokens.
Key Facts
| Details | |
|---|---|
| Model | Nanbeige4.3-33B (plus Nanbeige4.3-33B-Base) |
| Lab | Nanbeige LLM Lab / BOSS Zhipin (南北阁) |
| Size / type | 33B dense (BF16 on the card) |
| License | Apache 2.0 (Hugging Face cardData / tags) |
| Context | Up to 524,288 tokens (512K) |
| Role | Flagship teacher for Nanbeige4.3; planned distillation to smaller models (e.g. Nanbeige4.3-3B) |
| Focus | Agentic coding and professional cowork |
| Self-reported leads | Lab table vs Qwen3.8-27B, Muse-Glimmer-30B, Gemma4-31B; e.g. Terminal-Bench 2.1 80.9 vs Qwen 73.0 (publisher figures) |
| HF created | About October 24, 2026 (API createdAt); covered by MarkTechPost on October 25, 2026 |
Background
The model card frames Nanbeige4.3 as a teacher-first strategy: ship a strong ~30B agent, then distill into lighter models for deployment. Quickstart recipes cover Transformers, SGLang, and vLLM, with recommended settings that differ for agent/tool use versus reasoning/chat. The Base page posts a separate self-reported table against Gemma4-31B-Base on MMLU-Pro, GPQA and related tasks.
MarkTechPost's October 25 report restates the Apache 2.0 release, 512K context, and the lab's claim that Nanbeige leads 17 of 20 rows in its own comparison. It also flags that some competitor scores in Nanbeige's table do not match those labs' own Hugging Face cards (for example GPQA Diamond for Gemma 4 31B, and swapped-looking SWE-Bench Pro figures for Muse-Glimmer versus Gemma). Treat publisher tables as provisional until independent harnesses reproduce them.
What We Don't Know Yet
- Independent third-party runs of Terminal-Bench, SWE-Bench Pro, DeepSWE and the cowork suites under a shared harness.
- Why several competitor numbers in Nanbeige's table diverge from those labs' own cards (harness differences versus copy errors).
- Release date and specs for the planned Nanbeige4.3-3B distillate.
- Hosted inference availability (MarkTechPost noted no Hugging Face Inference Provider host at publication).