The new flagship rolls out across kimi.com, Kimi Work, Kimi Code, and the API with native image understanding and a million-token context. Moonshot puts its scale around 2.8 trillion parameters and its overall intelligence just behind Claude Fable 5 and GPT-5.6 Sol — at a fraction of frontier closed-model pricing.
2026-07-16 · by Moe Ameen
Moonshot AI, the Beijing lab behind the Kimi series, began rolling out Kimi K3 — its new flagship frontier model — in mid-July 2026, serving it across kimi.com, the Kimi Work and Kimi Code apps, and the Kimi API. It is the successor to the K2 line and Moonshot's most capable model to date, positioned for reasoning, agentic work, and multimodal understanding.
Moonshot describes K3 as a very large model — its figures put the total around 2.8 trillion parameters, which the lab frames as the largest open-weight model released so far (early coverage cited numbers in the roughly 2.5–2.8 trillion range, so treat the exact count as approximate until the model card is public). The features everyone agrees on: a 1-million-token context window paired with a new attention design built for fast decoding over very long inputs, and native multimodal understanding, with visual input built into the model rather than bolted on. Early hands-on testers highlighted strong front-end/UI and 3D generation. At launch Moonshot listed API pricing around $3 per million input tokens and $15 per million output (cached input near $0.30), undercutting the frontier closed models it is compared to. Full open weights were expected to follow the hosted launch, with early reports pointing to on or around July 27, 2026.
On performance, Moonshot's own framing places K3's overall intelligence just behind Claude Fable 5 and GPT-5.6 Sol, and reporting around the launch said it was expected to match or beat Anthropic's Opus 4.8 on some benchmarks while trailing Fable 5, with especially strong visual output. Those are first-party claims and pre-release impressions, and the loudest comparison videos circulating on social media were community tests rather than published evaluations — so treat the numbers as provisional until independent leaderboards land. The broader story analysts drew: a Chinese lab shipping a bigger, cheaper, open frontier model narrows the assumed gap between Chinese and US models and pushes frontier intelligence further toward a commodity.
There are two ways to act on this launch today, and Kompozy powers both. The first is the obvious one for a creator: this is a news moment your audience is already searching, so publish your take on it this week. Drop a short brief — what K3 is, the 1M context, the open-weight angle, why cheap frontier models matter — into Kompozy and it becomes a same-day package: a captioned Persona Short or Clipped Short explaining the launch, a brand-exact Carousel comparing K3 to Fable 5 and GPT-5.6, a Quote Graphic on the standout stat, native Text Posts for X and LinkedIn, a Blog Article for the long-tail search, and an Email Newsletter to your list — all held to one voice by your Persona Brief and reframed to 9:16, 1:1, and 16:9, then scheduled and published across nine platforms plus blog and email from one queue.
The second is the strategic read, and it is the more durable one. K3 is another data point in frontier intelligence going cheap, open, and interchangeable — which quietly moves the advantage away from which model you can access and toward the production-and-distribution layer wrapped around it. That layer is exactly what Kompozy is. You don't pick or operate a raw model; Kompozy runs its own managed Claude and OpenAI models and hands you outcomes — 18 content formats, persona and avatar video a model can't render, brand-voice governance, and full autopilot publishing. So a launch like K3's isn't a threat to that workflow; it is the kind of raw intelligence the engine stands on while you keep the part that actually compounds: a recognizable brand, shipped everywhere, on schedule.
Kimi K3 is Moonshot AI's new flagship frontier model, rolling out in mid-July 2026 across kimi.com, Kimi Work, Kimi Code, and the Kimi API. It is a very large, natively multimodal model with a 1-million-token context window, described by Moonshot as around 2.8 trillion parameters and the largest open-weight model to date, built for reasoning, agentic work, and image understanding.
Moonshot framed K3 as the largest open-weight model to date and said weights would follow the hosted launch (early reports pointed to on or around July 27, 2026). At launch it listed API pricing around $3 per million input tokens and $15 per million output, with cached input near $0.30 — cheaper than the frontier closed models it is compared to. Confirm current specs on Moonshot's own pages.
Moonshot's own framing places K3's overall intelligence just behind Claude Fable 5 and GPT-5.6 Sol, and launch reporting said it was expected to match or beat Opus 4.8 on some benchmarks while trailing Fable 5, with strong visual output. Those are first-party claims and pre-release community impressions — treat them as provisional until independent leaderboards confirm them.
Not finished content. K3 drafts text, reasons over long inputs, and understands images, but it renders no captioned video, branded graphics, or scheduled posts. To act on the launch as a creator — an explainer short, a comparison carousel, a blog, and platform-native posts — you run a draft through a content engine like Kompozy, which generates 18 formats and publishes across nine platforms plus blog and email.