Kimi K3 Review

The 2.8 trillion parameter count is the headline and the least interesting thing about K3. What matters is that Moonshot published the weights. Almost nobody outside a handful of companies can serve a model this size, so the release works less as a gift to hobbyists than as a statement about where the frontier sits and who is allowed to stand on it. The architecture is a sparse mixture of experts, 896 experts with sixteen active per token, a one million token context window, and understanding of images and video built in rather than bolted on through an adapter.

The lineage explains the ambition. K2 in July 2025 was the first trillion parameter open release. K2.5 in January 2026 added vision and Agent Swarm, which coordinates up to a hundred sub agents at once. K2.6 in April pushed that to three hundred agents across four thousand coordinated steps. K3 is the point at which the line stops being interesting for an open model and starts being interesting without the qualifier, and long horizon coding and knowledge work are where it earns the grades below.

The catch is the licence and the plumbing. Any provider above $20 million of annual revenue serving it externally has to negotiate a contract with Moonshot, which is open weights with a toll booth attached, and self hosting is theoretical for all but a few. In practice people reach it through the app or the API, where pricing lands near Claude Sonnet and well under the American flagships. A free tier with rate limits sits below four paid tiers named after tempo markings.