A clearer view of AI.
THE INDEPENDENT AI.

FRONTIER LAB / LLM · NICHE

Nanbeige (Nanbeige LLM Lab)

Chinese AI lab releasing open-weight models.

Visit github.com ↗@nanbeige on X ↗

What Nanbeige (Nanbeige LLM Lab) does

Chinese AI lab releasing open-weight models. Nanbeige4.3-33B (released 25 Oct 2026) is a 33B dense 'teacher' model for agentic coding and office ('cowork') tasks, Apache 2.0, with a 524,288-token context and XML tool calling; a smaller Nanbeige4.3-3B distilled model is promised. Model card: huggingface.co

Good for

  • Test it as a private, self-hosted coding or back-office agent model where client data must stay on your own servers
  • Compare it with Qwen or Gemma models of similar size on your own tasks before picking an open model for a local agent

Try this in 10 minutes

  1. Test it as a private, self-hosted coding or back-office agent model where client data must stay on your own servers. Use the free version: pick one real example from your own business and try just that.
  2. Compare it with Qwen or Gemma models of similar size on your own tasks before picking an open model for a local agent. Put the result next to how you do it today. Is it faster, cheaper or better?
  3. Sign up (free plan or trial if available) and try it on one real task from this week.

Mini-tasks based on common small-business uses for this tool.

Honest take

Some claims not independently checked

  • All benchmark numbers are the lab's own, with no independent reproduction yet; Qwen3.8-27B still edges it on SWE-Bench Pro and Gemma4-31B on GPQA Diamond.
  • Not laptop-friendly at full precision.

What we couldn’t verify

  • Fact vs hype: weights and model card are real on Hugging Face (Nanbeige/Nanbeige4.3-33B).

Pricing

Free open weights on Hugging Face (Apache 2.0). You pay for your own hardware or hosting; the full BF16 model is ~66 GB, and the lab's serving example uses 4 GPUs. Community GGUF quantisations exist for local use.

Prices as listed by the maker when we last checked (2026-10-26); usually in US dollars. Check the website for current prices.

Latest update

  • 2026-10-25: Released Nanbeige4.3-33B on Hugging Face; the lab says it beats Qwen3.8-27B and Gemma4-31B on most agentic benchmarks (self-reported). Flagged by @kimmonismus.

Is Nanbeige (Nanbeige LLM Lab) worth it? Quick answers

Is Nanbeige (Nanbeige LLM Lab) worth it?

Nanbeige (Nanbeige LLM Lab) is a niche pick: useful for specific cases, but most small businesses can skip it for now. It is free or open source, so the main cost is your time to set it up. Some of the maker's claims have not been independently checked.

How much does Nanbeige (Nanbeige LLM Lab) cost?

Free open weights on Hugging Face (Apache 2.0). You pay for your own hardware or hosting; the full BF16 model is ~66 GB, and the lab's serving example uses 4 GPUs. Community GGUF quantisations exist for local use.

What is Nanbeige (Nanbeige LLM Lab) used for?

Chinese AI lab releasing open-weight models. Typical uses: test it as a private, self-hosted coding or back-office agent model where client data must stay on your own servers; compare it with Qwen or Gemma models of similar size on your own tasks before picking an open model for a local agent.