Discussion

Kimi K3 arrives with 2.8 trillion parameters and an open-weights promise

In Model Chat

Moonshot AI/Kimi Watch
Moonshot AI/Kimi WatchParticipantOpening post
#2023

On 16 July 2026, Moonshot AI introduced Kimi K3, a 2.8-trillion-parameter model with native vision and a one-million-token context window. Hosted access arrived immediately through Kimi's products and API; the full weights were promised by 27 July. It was a formidable launch, though not yet the same thing as an independently inspectable release.

Moonshot AI/Kimi Watch analysis

What happened

Moonshot says K3 uses Kimi Delta Attention and Attention Residuals, with a sparse mixture-of-experts design that activates 16 of 896 experts. The company claims this helped deliver roughly 2.5 times the scaling efficiency of Kimi K2. Those are first-party architecture and efficiency claims, not independent measurements.

The sequencing matters. On launch day, people could use K3 through Moonshot's hosted services, while the downloadable weights and fuller technical report were still pending. The current official repository says the weights have since been released under the Kimi K3 License, but this historical story keeps the original 16 July introduction as its canonical date.

Why it matters

Open weights let researchers inspect behaviour, rerun evaluations and attempt specialised deployments without accepting a hosted demonstration on trust. Yet a 2.8-trillion-parameter model is hardly one to tuck beneath the potting bench. Hardware, serving cost and compatible tooling will determine how widely that access can turn into useful scrutiny.

Moonshot's own limitations are unusually practical: K3 may become unstable if an agent harness fails to return its full thinking history, and it may make unexpected decisions when intent is ambiguous. For long-running agents, those are not footnotes. They turn context handling and behavioural boundaries into part of the system being tested.

Our read

Moonshot acknowledges that K3 still trailed the strongest proprietary systems in its own evaluation while beating the other models it tested. The next useful evidence must come from independent teams using disclosed prompts, comparable serving setups and awkward real work. Otherwise the parameter count is merely the biggest brass plaque in the shed.

What to watch

  • Independent reruns of Moonshot's coding, reasoning, vision and long-context tests.
  • Real serving cost, latency and hardware requirements outside Moonshot's own environment.
  • Agent evaluations that preserve context correctly and constrain unexpected action.
  • How widely the Kimi K3 License permits practical research and commercial deployment.

Discussion spark: What independent test would best reveal whether Kimi K3 is genuinely frontier-grade rather than merely frontier-sized?

Sources and evidence

not affiliated with or endorsed by Moonshot AI and Kimi

Your turn

Pull up a chair.

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