Kimi K3.1 Pro Weights Are Out: 2.8T Parameters, Who Should Care
In short
- Moonshot AI released the open weights of Kimi K3.1 Pro on 4 November: 2.8T parameters, 104B active, 1M context.
- Weights alone are about 1.4 TB at 4 bits, so almost nobody should self-host it.
- API list price is $2 in / $10 out per million tokens, about €25 a month for a busy agent.
Moonshot AI put the full weights of Kimi K3.1 Pro on Hugging Face on 4 November. A 2.8-trillion-parameter model, free to download. Does that change anything for you? Yes, but not the way the headlines say.
What happened
Moonshot launched K3.1 Pro on 23 October as an API model and promised the weights by 4 November. They kept the promise. It is a mixture-of-experts model: 2.8 trillion parameters in total, 104 billion active per token, a 1-million-token context window, quantised to MXFP4. The model card recommends vLLM and SGLang to run it.
The licence is a custom “Kimi K3.1 Pro License”, not Apache. Read it before you build a product on it. The benchmarks come from Moonshot itself: Terminal-Bench 4.0 at 65.7, GPQA-Diamond at 94.2, BrowseComp at 92.1. The model card’s own table shows it winning some tests against Claude, GPT, DeepSeek and GLM and losing others. Treat the numbers as a sales sheet.
The API list price is $2 per million input tokens and $10 per million output tokens, with cached input at $0.20. Check platform.kimi.ai before you plan a budget.
How to use it
Most people should not download it. Use it through the API, or through a hosting provider that picks up the weights over the next weeks. Test it on one real task from your business, not on a benchmark.
Do the math: at 4 bits, 2.8 trillion parameters need about 1.4 TB just for the weights (2.8T x 0.5 byte, my estimate). That is roughly eighteen 80 GB GPUs before a single token of context. Now the API side. An agent that reads 200,000 tokens and writes 50,000 per day costs $0.40 + $0.50 = $0.90. That is about €25 a month. Self-hosting only wins if you run it around the clock at scale.
What it means
The real news is the exit door. Once the weights are public, any provider can host the model and compete on price. If Moonshot raises prices, you move. Moonshot already closed K2 to new users, so this matters.
Who loses? Closed labs that charge a premium for models that open ones now match on several tests. Who wins? Anyone who pays for tokens every month.
Ivan’s Take
I would not run this on my own server. 1.4 TB is not a hobby. I would test it through the API on one real job from my business, compare the bill with what I pay now, and keep it as a fallback. Open weights are worth it for the exit door alone.
Source: Moonshot AI
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
