Discussion

Moonshot’s Kimi K3.1 Pro pushes towards the open-weight frontier, but the weights come later

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#8866

Moonshot has launched Kimi K3.1 Pro through its website, app and API, with independent evaluations placing it close to leading proprietary models on several tests. The catch is in the name: the model weights are promised for 4 November, not available at launch.

Watch Desk analysis

What happened

Kimi K3.1 Pro is a 2.8-trillion-parameter mixture-of-experts model, with 104 billion active parameters and a one-million-token context window, according to Moonshot’s launch details as summarised by Latent Space. Standard API prices are $2 per million input tokens and $10 per million output tokens, a third lower than Kimi K3’s listed rates. Moonshot says its Highspeed version generates output two to three times faster, at twice the standard price.

The release also includes Hosted Agents, a managed environment for long-running work and multi-agent collaboration. The weights are due on 4 November; the available account does not establish their eventual licence or redistribution terms, so “open-weight” is the safe description for now.

Why it matters

The evaluations suggest a meaningful step up, particularly on agentic tasks. Artificial Analysis placed K3.1 Pro fourth on its Intelligence Index, while Arena ranked it eighth in its WebDev coding category, up from seventeenth for Kimi K3. Those are results from different evaluations, not a universal verdict that it beats every closed model.

Cheaper tokens do not automatically mean cheaper completed work. Artificial Analysis found the model’s Intelligence Index task cost roughly unchanged from K3, as K3.1 Pro used more output tokens per task. Its reported Arena WebDev median task cost was lower than K3’s, while the Cyber Index task cost was higher. The bill depends on the job, not just the price list.

Our read

This looks like a serious frontier contender, and the forthcoming weights could make the comparison more consequential. But the launch-day distinction matters: people can use the hosted model and API now; they cannot yet run the weights themselves. Nor does 104 billion active parameters mean a 104-billion-parameter model fits on a typical local machine: the total expert pool and serving requirements matter too.

What to watch

  • Whether the weights arrive on 4 November, and under what licence.
  • How independent evaluations change once researchers can test the released weights directly.
  • Whether real-world task costs match the benchmark results, especially for long-running agents.
  • Whether Highspeed’s extra cost buys enough time saved to justify paying twice as much.

Discussion spark: When the weights arrive, should openness be judged mainly by access to the model, or by whether people can realistically run and adapt it?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.