Moonshot open-sources Kimi K3.1 Pro and sells the agent harness
Moonshot handed out the thing labs usually keep locked up — and started charging for everything around it. Open weights plus a managed runtime. That's how infrastructure companies behave. Labs don't move like this.
What shipped
Kimi K3.1 Pro weights are on Hugging Face. Multimodal, mixture-of-experts: 2.8 trillion total parameters, roughly 104 billion activated per inference, context around one million tokens. Deploy it on your own metal.
The numbers explain why a flagship can be given away at all. MoE wakes only part of the experts, so the inference bill stays sane. Strip that architecture out and nobody could afford to run it.
What's for sale
Kimi Hosted Agents is the wrapper: sandboxes, audit trails, an Inference Hook, eight finance templates, running around the clock. Teams skip the runtime work.
The logic is blunt. Open weights push the model worldwide for free; the harness is where money is made. Open weights do the marketing, the harness does the revenue. Moonshot isn't picking between two business models — it runs both.
Why this beats another benchmark
Western labs still argue about whether to open a flagship. That argument has run for years. Moonshot shipped the decision as a product.
Second, agent harnesses are turning into a commodity. A year ago "build an agent" meant stitching a pipeline by hand: sandbox, logs, retries, safe tool access. Now it's a subscription. The barrier drops, and whoever has cheap inference plus a ready runtime wins. Moonshot has both.
Third, a million-token context isn't about dumping a document library. It's about an agent that holds the whole session and doesn't lose the thread. Without long context, hosted agents stay a toy. With it, they're a tool.
The risk: managed agents mean someone else's infrastructure holds your data. Auditing and sandboxing aren't decoration — they're the deal. Enterprises ask first, and Moonshot answers with templates and logs, not promises.
Bottom line: Moonshot doesn't sell a model. It sells the place where the model lives.
💬 What to argue about
- Are open flagship weights a strategy or a forced move? If the model is free, where's the moat once any cloud provider can copy the harness?
- Managed agents vs self-hosted: what has to change in inference pricing or regulation before teams pull agents back onto their own hardware?
- Is a million-token context a real need or a spec race? Name a task where it pays off and one where it's dead weight.
Prediction: Within months, at least one Western provider will ship a comparable managed agent runtime — open weights made the harness the real battlefield.
Disclaimer: not officially confirmed — the cluster has no Moonshot publication, only industry outlets.
🤖 AI summary
- Date: 2026-11-05
- Source: technode.com
- Tags: Moonshot, Kimi, Agents
- Key facts:
- Moonshot handed out the thing labs usually keep locked up — and started charging for everything around it.
- Open weights plus a managed runtime.
- That's how infrastructure companies behave.
- Labs don't move like this.
- Kimi K3.1 Pro weights are on Hugging Face.
- Multimodal, mixture-of-experts: 2.8 trillion total parameters, roughly 104 billion activated per inference, context around one million tokens.
- Deploy it on your own metal.
- The numbers explain why a flagship can be given away at all.
Source: technode.com · post in Telegram