Moonshot preps Kimi K3 launch at 2-3 trillion parameters

The most striking thing about the leaks around Moonshot's next model isn't the parameter count, it's the timing. According to Financial Times reporting picked up by other outlets this week, the Chinese startup is preparing to launch Kimi K3 in the coming days, with people familiar with the plans describing a model in the two to three trillion parameter range.

If those numbers hold, K3 would sit among the largest open-weight models ever released, and it arrives in Moonshot's mixture-of-experts lineage rather than as a straight dense scale-up. The reporting also says Moonshot expects the model to outperform Anthropic's Claude Opus on several widely recognized benchmarks, and leaked descriptions of a one million token context window suggest the design is aimed squarely at long-horizon coding and agent workloads rather than chat.

The launch is being lined up around WAIC 2026, the conference slot Chinese labs have used before to make a statement. That framing matters more than any single number. A trillion-parameter open-weight release from a Chinese lab, timed to a domestic showcase, is a signal to developers everywhere that the frontier is no longer only in San Francisco, and to policymakers that the release cadence they hoped to slow has kept moving.

The honest caveat is that almost none of this is confirmed. Moonshot has published no specs, no model card, and no benchmark suite for K3, and the FT reporting rests on unnamed people familiar with the plans. Community testers with beta access have already produced conflicting impressions. Treat the parameter count, the launch window, and the Claude comparison as reported, not settled, until the release actually lands.

What is worth watching, if K3 does ship as described, is who reaches for it. Small teams priced out of Anthropic and OpenAI's premium tiers, developers building long-context agent workflows, and buyers outside the US shopping for a frontier open-weight option would all benefit. The gap between the leak narrative and the eventual model card will tell you how much of that upside is real.