Moonshot is positioning K3 not just as a giant model, but as a distribution and product strategy for long-horizon agent work
Moonshot AI's official Kimi K3 post frames the model as the world's first open 3T-class model: roughly 2.8 trillion parameters, a 1-million-token context window, native vision, and product integration across Kimi.com, Kimi Work, Kimi Code, and the API. The key takeaway is not simply size, but how Moonshot wants K3 to be used in real workflows.
Scale
2.8T parameters; described as the first open 3T-class model.
Context
1M-token context window with native vision capabilities.
Architecture
Built on KDA, AttnRes, and Stable LatentMoE; activates 16 of 896 experts.
Rollout
Available across Kimi products now, with full weights scheduled by July 27.
The official post is notable for its tone. Moonshot does not declare that Kimi K3 has surpassed every proprietary rival. Instead, it explicitly says the overall experience still trails Claude Fable 5 and GPT 5.6 Sol, while arguing that K3 has reached frontier-level performance across the company's evaluation suite. The strategic claim is therefore more specific: an open giant model is now close enough to matter at the frontier.
Architecturally, Moonshot highlights Kimi Delta Attention, Attention Residuals, and Stable LatentMoE, claiming roughly 2.5x better scaling efficiency versus Kimi K2. The post also says K3 activates 16 experts out of 896. That matters because the announcement is not just about raw parameter count; it is about making a 2.8T model operationally usable.
The strongest emphasis is on long-horizon agent work. The blog showcases GPU kernel optimization, a MiniTriton compiler, game development, chip design, scientific coding, interactive research dashboards, and video editing. In other words, Moonshot is selling K3 less as a chatbot and more as an agentic work platform for sustained engineering and knowledge tasks.
The commercialization details are unusually concrete. Kimi K3 is available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with official pricing of $0.30/MTok for cache-hit input, $3.00/MTok for cache-miss input, and $15.00/MTok for output. Combined with the promised weight release, that makes the post as much a go-to-market document as a model launch note.
Moonshot also discloses limitations: sensitivity to preserved thinking history, a tendency toward excessive proactiveness in ambiguous tasks, and an acknowledged user-experience gap versus the strongest proprietary systems. That level of caveat is itself important, because it suggests the company expects K3 to be implemented in real environments, not merely benchmarked from afar.
Why this matters
- It reframes the competition around deployment and agent workflows, not just leaderboard scores.
- It pairs open-weight ambition with concrete product and pricing distribution.
- It suggests that the next phase of the model race will hinge on ecosystem reach as much as raw capability.