Kimi K3 released: 2.8 trillion parameters, native multi-modality, million token context

On July 16, 2026, Moonshot AI released Kimi K3 (the flagship model of the Kimi series), which is officially positioned as the "new frontier of intelligence." Three key numbers define its ambitions: 2.8 trillion parameters, native multi-modality, and million (1M) token context - designed for long-range programming, knowledge work and deep reasoning. The PerceptionBench evaluation benchmark was also released on the same day, providing a quantifiable yardstick for "multimodal perception capabilities".

Capability relay from K2.6 to K3

Kimi K3 didn’t come out of nowhere. The previously released Kimi K2.6 on April 20 has paved the way for a new generation of capabilities - a noteworthy detail is that the Dark Side of the Moon official website page itself was co-created by Kimi K2.6. This approach of "using your own models for your own products" is not only a verification of capabilities, but also a self-reinforcement of the product narrative.

From chat assistant to complete agent matrix

Kimi's product form has gone far beyond "chat assistant": it has expanded to a complete agent product matrix covering Agent, Agent cluster (agent-swarm), one-click website building, documents, PPT, forms, in-depth research (Deep Research) and Kimi Claw, and launched Kimi Code and open platform. K3's ultra-long context and native multi-modality provide the foundation for long-term tasks in this matrix (long document processing, long code, in-depth research).

Representative sample of scale tangential value realization

At the industry level, according to 36Kr reports, Kimi ARR has exceeded US$200 million - this is one of the representative samples of China's large model "from scale competition to value realization". At a time when large domestic models are generally still burning money to gain scale, Dark Side of the Moon took the lead in telling the revenue story. The release of K3 further consolidates its leading position among domestic long-context models.

From an industry perspective, Kimi K3's "2.8 trillion parameters + 1M context + native multi-modality" combination marks that the focus of competition for domestic models is shifting from "can it be used" to "can it withstand long tasks?" Ultra-long context is not a show-off, but a basic capability for real paid scenarios such as in-depth research, long documents, and agents - it is these "high-value long tasks" that K3 is betting on.

Several directions worth tracking in the future:

1. K3’s contextual cost-effectiveness: Whether the power of 1M token is affordable in terms of real cost.
2. PerceptionBench’s credibility: Whether the multi-modal evaluation benchmark will become an industry benchmark.
3. Agent-swarm maturity: The ability of the Agent cluster to orchestrate complex tasks.
4. ARR sustainability: After US$200 million in ARR, can the paid scenario continue to expand.