Moonshot launches Kimi K3, API-only for now

Moonshot AI’s Kimi K3 went live on the company’s Kimi API Platform on July 16, after two days of leaks and unofficial testing. Moonshot’s own documentation lists K3 at 2.8 trillion parameters with a 1,048,576-token (roughly 1M) context window, built on a hybrid linear-attention design the company calls Kimi Delta Attention, paired with what it terms “Attention Residuals.” The docs describe always-on reasoning with a configurable effort setting, tool calling, JSON-schema output, and automatic prompt caching. API pricing is $3 per million input tokens on a cache miss, $0.30 on a cache hit, and $15 per million output tokens.

Two things stand out. Moonshot has not published an active-parameter count or confirmed a mixture-of-experts label for K3, unlike its K2-family models, which disclosed roughly 32 billion active parameters out of a trillion total. And K3 shipped API-only: there is no Hugging Face model card, license, or downloadable weights as of publication, a departure from K2, K2.5, K2.6, and K2.7-Code, all of which released open-weight. Moonshot’s own homepage still lists K2.6 as its headline model, with no blog post announcing K3.

Benchmark claims that K3 beats GPT-5.5 or Claude Opus on specific coding tests are circulating from testers on evaluation arenas and social media, not from a Moonshot-published benchmark suite. Treat them as unverified for now.

For builders, the practical read is to wait for Moonshot’s own spec sheet, and if you rely on open-weight Kimi models for self-hosting, not to assume K3 will follow that path. Thinking Machines’ Inkling shipped the opposite way, open weights from day one, the same week; see open models good enough for operators for how to weigh a new entrant before switching anything.