Moonshot AI announced Kimi K3, its new flagship model, on 16 July 2026. Moonshot frames it as the first open model in the 3-trillion-parameter class. One detail matters before any of the excitement: the weights are not downloadable yet. The model is live today in Moonshot’s own app and API, and the open weights are scheduled to arrive on Hugging Face on 27 July 2026.

A glowing orange filament sphere suspended inside a dark concrete chamber, standing in for the core of a large neural network.
Kimi K3 is a mixture-of-experts model: 2.8 trillion parameters in total, but only a small fraction light up for any single token.

What happened

Moonshot AI, the Chinese lab behind the Kimi assistant, published the Kimi K3 announcement on its official blog. The headline numbers, as stated by Moonshot:

  • 2.8 trillion total parameters, in a mixture-of-experts design that activates 16 of 896 experts per token (Moonshot calls the routing scheme “Stable LatentMoE”). A mixture-of-experts model keeps inference cost far below what the total parameter count suggests, because most of the network stays idle on any given token.
  • A new attention architecture, which Moonshot names Kimi Delta Attention and Attention Residuals, alongside Gated MLA and low-precision MXFP4 weights.
  • A 1 million token context window and native vision (text, images, and video as input).
  • Reasoning on by default, with lower- and higher-effort thinking modes promised in later updates.

The model is available now in the Kimi app (iOS, Android, HarmonyOS), the Kimi Work desktop app, the Kimi Code command-line tool, and the Kimi API. API pricing, per Moonshot’s platform, is about €0.28, €2.75, and €13.80 per million tokens (USD 0.30 cache-hit input, 3.00 input, 15.00 output). Open weights are listed on Hugging Face as an upcoming release dated 27 July 2026, so as of today the model is announced-open, not yet downloadable.

Moonshot published benchmark comparisons against Claude Fable 5, Claude Opus 4.8, GPT-5.6 Sol, and GPT-5.5. Treat these as the lab’s own claims: they include leading scores on some coding and terminal benchmarks, and, to Moonshot’s credit, an explicit admission of a “noticeable gap in user experience compared with Claude Fable 5 and GPT-5.6 Sol.” Moonshot has not stated the license for K3, so its openness terms are not yet known.

Why it matters for builders

The open-weight tier keeps climbing. If the weights land as scheduled, K3 joins DeepSeek-V4 and Qwen’s open family as a frontier-scale model you could, in principle, self-host. That matters when you need data residency, offline deployment, or freedom from a single vendor’s pricing.

Two cautions keep this honest. First, an announced model is not a shipped one: until the 27 July weights and the license appear, K3 is something you can call through Moonshot’s API but not run yourself, and “open” cannot be confirmed. Second, the benchmark lead is Moonshot’s own measurement, and the lab itself concedes the day-to-day experience still trails the top closed models. For a build decision, wait for the weights, read the license, and run your own evaluations rather than trusting any single leaderboard. See the 2026 LLM landscape for how the families compare.

Sources

Further reading