Moonshot AI's new flagship frontier model — a very large, long-context, natively multimodal model that reads images and reasons over million-token inputs, positioned as the largest open-weight model from China.
Last verified · 2026-07-16 · by Moe Ameen
Kimi K3 is the newest flagship model from Moonshot AI, the Beijing lab behind the Kimi series. Moonshot began rolling it out in mid-July 2026 across kimi.com, its Kimi Work and Kimi Code products, and the Kimi API, positioning it as a frontier-class model for reasoning, agentic work, and multimodal understanding. It is the successor to the K2 line (including the K2.7 Code coding model) and Moonshot's most capable model to date.
On scale, Moonshot describes K3 as a very large mixture-of-experts model — its figures put the total around 2.8 trillion parameters, which the lab says makes it 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. Like earlier Kimi models it is expected to activate only a fraction of those parameters per token, which is how a model at this scale stays cheaper to serve than a dense one of the same size. The headline capability everyone agrees on is a 1-million-token context window, paired with a new attention design Moonshot built for fast decoding over very long inputs.
K3 is natively multimodal: visual understanding is built into the model rather than bolted on as a separate module, so it can read images and screenshots as first-class input alongside text. Early hands-on testers highlighted unusually strong front-end/UI and 3D generation from prompts. At launch Moonshot listed API pricing around $3 per million input tokens and $15 per million output (with cached input near $0.30), undercutting the frontier closed models it is compared to. Full open weights were expected to follow the hosted launch — early reports pointed to on or around July 27, 2026 — likely under a permissive, modified-MIT-style license as with prior Kimi releases.
On performance, Moonshot's own framing places K3's overall intelligence just behind Claude Fable 5 and GPT-5.6 Sol, and early third-party impressions suggested it matches or beats 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 — worth noting, worth waiting on independent leaderboards to confirm. The honest framing: K3 is a powerful reasoning-and-understanding engine you prompt for text and analysis. It is not a content studio — it drafts and reasons, but it renders no finished video, no branded image, no scheduled post.
Think of Kimi K3 as the drafting and reasoning brain and Kompozy as the studio and distribution floor. K3 is exceptional at the thinking half — you can paste an hour-long transcript, a research folder, or your last fifty posts into its million-token context and get back a sharp outline, a batch of angles, a script, or a newsletter draft. What it hands you is words and analysis. It does not render a captioned vertical short, a face-locked avatar video, a brand-exact carousel, a quote card, or a scheduled post — and that is the entire second half of a creator's job. That is where Kompozy picks it up: paste a K3 draft or brief in as your source, and Kompozy generates the finished media across 18 formats — Persona Shorts and HeyGen avatar video, Clipped Shorts, Carousels and Quote Graphics via brand-exact HyperFrames, Photo Posts, a Blog Article and an Email Newsletter — every output held to one voice by your Persona Brief instead of reading like raw model output.
The clean division of labor is why the pairing works: K3 is a raw model you prompt (through a chat window, Kimi Work, or the API), while Kompozy is the generation-and-publishing engine that turns intent into finished, on-brand assets and fans them across nine social platforms plus blog and email from a single queue, with Autopilot and a per-post review pipeline. You do not even have to operate K3 yourself — Kompozy runs its own managed Claude and OpenAI models for generation, so the practical workflow is: use K3 when you want its long-context reasoning to shape the idea, then let Kompozy produce the video, images, and posts and ship them. K3 sharpens the plan; Kompozy makes the plan into a published week.
Kimi K3 is Moonshot AI's newest flagship 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 total parameters and the largest open-weight model to date. It is built for reasoning, agentic work, and image understanding.
Yes for input. Visual understanding is built into the model, so K3 reads images and screenshots as first-class input alongside text, and early testers highlighted strong front-end/UI and 3D generation. Its core generated output is text and code — it does not produce finished social video, branded graphics, or scheduled posts.
At launch Moonshot 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. Moonshot said it would publish open weights after the hosted launch (early reports pointed to on or around July 27, 2026), likely under a permissive modified-MIT-style license as with prior Kimi releases. Confirm current specs on Moonshot's own pages.
No. K3 drafts text, reasons over long inputs, and understands images, but it renders no captioned video, brand-consistent images, carousels, or scheduled posts. To turn a K3 draft into finished, on-brand content, pair it with a content engine like Kompozy, which generates 18 formats and publishes to nine social platforms plus blog and email.
Moonshot's own framing places K3's overall intelligence just behind Claude Fable 5 and GPT-5.6 Sol, and early impressions suggested it matches or beats Opus 4.8 on some benchmarks while trailing Fable 5, with especially strong visual output. Those are first-party claims and pre-release impressions — wait for independent leaderboards before treating any single number as settled.