Moonshot AI will release full Kimi K3 weights on July 27, 2026 — the largest open-weight frontier model yet, with hosted API access live today via Kimi Code and OpenRouter.
Moonshot AI is running a staged release: Kimi K3 is already live on kimi.com, Kimi Work 3.1+, Kimi Code, and the Kimi API at $0.30/$3.00/$15.00 per million tokens (cache-hit input, cache-miss input, output). OpenRouter exposes the model as moonshotai/kimi-k3 with the full 1M context, tool calls, and structured outputs. Full open weights arrive on July 27, 2026, which Moonshot says will be accompanied by a technical report covering architecture, training, and evaluation methodology.
Until the checkpoint drops, Artificial Analysis classifies Kimi K3 as proprietary — a distinction that matters for enterprises weighing data-sovereignty and auditability. Moonshot is also preparing Kimi Hosted Agent, an enterprise platform with agent harnesses, isolated sandboxes, and long-running environments. The disaggregated Mooncake inference stack reports cache hit rates above 90% on coding workloads, compressing effective input cost.
The open-weight release will be the largest ever at 2.8T parameters, giving global developers the option to fine-tune, self-host, or inspect a frontier-class model — a direct counterweight to the closed, government-gated releases from US labs this month.
Our Take
The hosted-first, weights-later playbook lets Moonshot capture API revenue while building developer goodwill before the open-source drop. For corp-dev teams, the July 27 release date is a procurement checkpoint: self-host economics only become calculable once weights and eval harness details are public. Until then, treat Kimi K3 as a competitive hosted option, not yet an infrastructure bet.