Nanbeige4.3-33B: A Free Agent Model You Can Run Yourself
In short
- Nanbeige released Nanbeige4.3-33B on Hugging Face on October 24 under Apache 2.0
- 33B parameters, up to 512K tokens of context, about 66 GB in BF16 or about 20 GB at 4-bit
- Its own scores beat Qwen3.8-27B on agent tasks, and a $400 a month Opus bill becomes $0 in tokens if you host it yourself
A free model built for agents just landed, and the weights are yours. Nanbeige published Nanbeige4.3-33B on October 24. The Apache 2.0 license means you can use it in a business, change it, and keep your data at home.
What happened
The model has about 33 billion parameters and a context window of up to 512K tokens. It is aimed at agentic coding and office-style work: long tasks where an AI calls tools, edits files and checks its own results.
Nanbeige published a table against Qwen3.8-27B. Here are the numbers.
- Terminal-Bench 2.1: 80.9 against 73.0
- DeepSWE: 65.5 against 42.5
- Automation Bench: 54.2 against 38.3
- SWE-Bench Pro: 60.3 against 61.7 (Qwen ahead)
- GDPval: 84.8 against 86.4 (Qwen ahead)
Now the honest part. Every score comes from the lab itself, run on its own setup, and nobody has reproduced them yet. On GPQA Diamond, Gemma4-31B scores higher (91.3 against 89.5). The gains are biggest on long multi-step tasks, which is exactly where agents live.
How to use it
Start with hardware. The weights take about 66 GB in BF16, so one 80 GB GPU gets tight once long context eats memory. At 4-bit it needs about 20 GB. That puts it within reach of a single high-end consumer card, but that is my guess, so test it first.
Ways to run it:
- Docker Model Runner: `docker model run hf.co/Nanbeige/Nanbeige4.3-33B`
- vLLM or SGLang for serving it to your team.
- Community quantized versions for llama.cpp, Ollama and LM Studio.
The model card suggests temperature 1.0 and top_p 0.95 for agent and tool tasks, and 0.6 with 0.95 for reasoning and chat. For tool calling it recommends the XML format.
Where to try it first: a private agent that reads your documents, triage for support tickets, or code refactors on a repo you cannot send to a cloud API.
Do the math: take a heavy agent workload of 50 million input tokens and 10 million output tokens a month (my assumption). On Claude Opus 5.5, at $4 and $20 per million tokens, that is $200 plus $200, so $400 a month and $4,800 a year. Self-hosted, the token bill is $0, but you pay for the GPU or a rented server. Against GPT-6 Luna at $0.10 and $0.50, the same load costs $10. So don’t switch for price alone if you can use a cheap hosted model. Switch for privacy and no lock-in.
What it means
Open models keep closing the gap on agent work, and hosted prices keep falling. The real choice is control against convenience.
Who wins? Teams with sensitive data and a spare GPU. Who waits? Everyone else, until independent tests confirm the numbers.
Ivan’s Take
I like free weights and I like keeping data at home. But a table the lab wrote itself is an ad, not a test. I would run Nanbeige on 20 of my own real tasks first, compare it with a cheap hosted model, and only then decide. If it holds up, an agent that never sends my data out is worth the GPU.
Source: Hugging Face
If it doesn’t save you time or money, you won’t see it here
- Facts, not hypeEvery price and number is checked against the official source
- Always the real priceYou see what it costs before you click
- I say when it’s weakEven when it’s OpenAI or Google
