Nanbeige4.3-33B: what it is and how to run it

Nanbeige4.3-33B is a free, open-weight AI model for coding agents and office-style automation. It went up on Hugging Face today under the Apache 2.0 license: a 33-billion-parameter dense model with a 512K-token context window that you can download, run on your own hardware and use commercially without paying per token.

It comes from Nanbeige LLM Lab, part of Kanzhun, the company that runs the Chinese job site BOSS Zhipin. @nanbeige announced it this morning and claimed "SOTA agentic performance among open models at ~30B parameters." According to that post, the lab built a strong teacher model first and plans to distill it into smaller ones, starting with a 3B version.

Why people on X noticed

A few hours later, @kimmonismus wrote that it beats similarly sized Qwen 3.8 27B and Gemma models on most benchmarks and called it a "really cool release." Size explains a lot of the attention. 33B is about as big as a model gets while still running on one good GPU or a well-equipped Mac.

The most useful post so far is from someone who tried it. @DogukanUrker got the model onto a 12GB RTX 3060 with a custom 2-bit quant of about 10GB. He saw roughly 20 tokens per second with a 24K context ceiling, and his verdict was "definitely not-usable for agentic work." For agent jobs he recommends a GPU with 16GB or more. He also pointed out that every GGUF published so far is 19GB or bigger, and that a plain 2-bit quant broke the model's reasoning until he kept a band of layers at 4-bit.

What the benchmarks say

Every number below is from Nanbeige's own model card. Nobody has reproduced them independently yet.

Nanbeige's biggest leads are on terminal and agent tasks. Qwen still edges it on SWE-Bench Pro and GDPval, and Gemma wins GPQA. The card also notes that Nanbeige ran its own model with preserve_thinking on and that the benchmarks used different harnesses, so I'd use these numbers to decide whether to test it, not as a verdict.

How to run Nanbeige4.3-33B

There are two realistic routes, depending on your hardware.

On a server with several GPUs

The official vLLM and SGLang examples on the model card split the model across four GPUs. Here is the vLLM version:

vllm serve Nanbeige/Nanbeige4.3-33B \
  --port 8000 \
  --tensor-parallel-size 4 \
  --gpu-memory-utilization 0.8 \
  --enable-auto-tool-choice \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

That gives you an OpenAI-compatible endpoint on port 8000, and most coding agents can point at that.

On one GPU or a Mac

The model card links to community quantizations that work with llama.cpp, Ollama and LM Studio. With the usable files starting around 19GB, I'd plan on a 24GB GPU or a Mac with plenty of unified memory. Smaller quants exist, but as Doğukan found, they cost you the context length and reliability that agents need.

Settings that matter

The card recommends:

• Temperature 1.0 and top_p 0.95 for agent and tool-use work, and 0.6 for plain chat and reasoning.
• Leaving thinking mode on, which is the default, unless you want quick answers. Set enable_thinking to false to turn it off.
• Turning preserve_thinking on for multi-turn tool use and code agents, and leaving it off for simple Q&A.

You can't call it through a hosted API yet. Hugging Face shows no inference provider serving it, so for now you either run it yourself or wait.

A first test worth running

Start with a task where you already know the right answer, on your own files. For agent work, something like:

Read the files in this folder. Find every script that writes to the customers table, list them, and write a one-paragraph summary of what each one changes.

If it gets that right, move up to a real bug.

My read

This is the second open coding model this week I'd bother downloading, after Recursive Flint, and it lands the same week as Kimi K3.1 Pro. Flint is the easier laptop install. Nanbeige needs more hardware, and if its numbers hold up, it's the stronger agent. It also fits what we wrote about open models handling most AI traffic: self-hosted models keep getting good enough for routine work.

For a business, the appeal is the usual one for self-hosting. Client files stay on your machines, and there's no per-token bill. Budget for real hardware, though, because a model limping along on a 12GB card won't save anyone time.

If you're working out which tasks could move to a local model and which still need a paid frontier one, that's the kind of question our free process audit is built to answer.