[AINews] Kimi 3.1 Pro 104BA-2.8T: Fable-5.1-Class Open Weights narrows Closed-AI lead to 1 month... and Pro Highspeed API with 2-3× speed
Latent Space swyx .ai ● Covered by 5 sources
Moonshot’s Kimi K3.1 Pro is out with open weights promised for Nov. 4. It’s a rare open model that’s pushing right up against the closed frontier.
Based on reporting by Latent Space, swyx .ai — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Moonshot has thrown Kimi K3.1 Pro into the ring, and the first read is simple: this is not a modest refresh. The company says the model has 2.8 trillion total parameters, 104 billion active parameters, and a 1-million-token context window. Open weights are promised for November 4, but for now the API is the only thing live.
The interesting part is less the splashy launch page than the benchmark shape underneath it. Artificial Analysis put K3.1 Pro fourth on its Intelligence Index with a score of 55, behind Opus 5.5 at 58 and just a step away from the next tier. Arena’s coding and preference tests were also strong: it landed #8 in Code Arena: WebDev with 1,692 points, and Arena says it improved across all categories enough to become its highest-ranked open-weight model once the weights arrive.
That doesn’t mean the model is a tidy all-rounder. The gains look especially agentic, not perfectly even across knowledge and science. And on the cyber side, Moonshot did something unusual: it intentionally removed cyber-related data during training and reinforcement learning. Even so, the model still scored 64% on Artificial Analysis’ Cyber Index, the highest publicly accessible result and second overall.
Pricing is where the story gets more complicated. Standard API rates are listed at $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached-input tokens, with a Highspeed version that Moonshot says runs 2–3x faster and costs twice as much. But cheaper tokens do not automatically mean cheaper work, and the reported task costs vary by benchmark: AI Index work came in about flat versus K3, while Arena’s WebDev median task cost was lower. The usual trap is assuming list price tells the whole story. It doesn’t.
Moonshot also paired the release with Kimi Hosted Agents, a managed execution setup for long-running work and multi-agent collaboration. That fits the broader pitch: not just a model, but a system for doing actual work. The open-weight claim matters because it moves the frontier conversation away from lab-only bragging rights and toward something people can inspect, run, and argue with in public.
My take — AI-written commentary, not fact-checked reporting
This is the part where the open-model crowd gets to be smug for once. A frontier-ish release with weights promised, strong coding, and a serious cyber score is exactly why closed shops hate transparency: it keeps ruining the story that only sealed systems can get this good. And yes, the industry will now spend another month pretending “open” means whatever helps the marketing deck that day.
Read more about this at: Latent Space