Kimi K3.1 Pro weights are out: 2.8T, ungated, and with a toll booth for big inference hosts
Moonshot kept its date. The Kimi K3.1 Pro weights went public on Nov 4, ungated on Hugging Face, at 2.8T total parameters with 104B activated. The model's numbers were old news from October. The license is the news.
It opens like MIT, then adds a clause aimed at one kind of business. If you or your affiliates run a Model as a Service business, and your aggregate revenue tops $20M over any consecutive 12 months, you need a separate agreement with Moonshot before using the model commercially at all. Note the trigger is total revenue, not revenue from this model. A cloud past that revenue line that sells inference on anything is probably in scope from day one.
The carve-outs are what make this workable for most people reading this. Internal use is exempt. End-user products with model capabilities embedded in specific features or harnesses aren't Model as a Service, and neither is relaying requests to models hosted by others. So you can put it inside your coding agent without a phone call. There's a badge rule too: over 100M monthly actives or over $20M monthly revenue, and "Kimi K3.1-Pro" must appear prominently in the UI.
So the pattern is settling. Mistral Large 4 shipped last week with a modified MIT license and its own MaaS clause. Now Moonshot does it too. Frontier open-weight labs have worked out who actually eats their lunch, and it's the inference host, not the app developer. Meanwhile Ricursive's Ripple-65B went out under plain Apache 2.0 the same day. Read the license before you build a business plan on the word open.
Practical reality: Moonshot's own table puts it at 65.7 on Terminal-Bench 4.0, against 66.4 for Opus 5.5 and 58.2 for GPT-6 Astra. Vendor numbers. And "self-hostable" at 2.8T means something very different from running a laptop model, so most people will still meet this through a certified inference partner. Two more on the shelf this week: Baidu released ERNIE 5.5 Flash Base, a 280B checkpoint taken at the end of pre-training with no mid- or post-training, explicitly as raw material for your own training, under its own custom license. Tencent launched Hy4 on Nov 4, but as of our Nov 5 check Hugging Face shows only the August preview, so don't count it as open yet. Weights: https://huggingface.co/moonshotai/Kimi-K3.1-Pro License: https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE ERNIE: https://ernie.baidu.com/release/ernie-5-5-flash-base/
← Back to all articles
It opens like MIT, then adds a clause aimed at one kind of business. If you or your affiliates run a Model as a Service business, and your aggregate revenue tops $20M over any consecutive 12 months, you need a separate agreement with Moonshot before using the model commercially at all. Note the trigger is total revenue, not revenue from this model. A cloud past that revenue line that sells inference on anything is probably in scope from day one.
The carve-outs are what make this workable for most people reading this. Internal use is exempt. End-user products with model capabilities embedded in specific features or harnesses aren't Model as a Service, and neither is relaying requests to models hosted by others. So you can put it inside your coding agent without a phone call. There's a badge rule too: over 100M monthly actives or over $20M monthly revenue, and "Kimi K3.1-Pro" must appear prominently in the UI.
So the pattern is settling. Mistral Large 4 shipped last week with a modified MIT license and its own MaaS clause. Now Moonshot does it too. Frontier open-weight labs have worked out who actually eats their lunch, and it's the inference host, not the app developer. Meanwhile Ricursive's Ripple-65B went out under plain Apache 2.0 the same day. Read the license before you build a business plan on the word open.
Practical reality: Moonshot's own table puts it at 65.7 on Terminal-Bench 4.0, against 66.4 for Opus 5.5 and 58.2 for GPT-6 Astra. Vendor numbers. And "self-hostable" at 2.8T means something very different from running a laptop model, so most people will still meet this through a certified inference partner. Two more on the shelf this week: Baidu released ERNIE 5.5 Flash Base, a 280B checkpoint taken at the end of pre-training with no mid- or post-training, explicitly as raw material for your own training, under its own custom license. Tencent launched Hy4 on Nov 4, but as of our Nov 5 check Hugging Face shows only the August preview, so don't count it as open yet. Weights: https://huggingface.co/moonshotai/Kimi-K3.1-Pro License: https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE ERNIE: https://ernie.baidu.com/release/ernie-5-5-flash-base/
Comments