Kimi K3 Just Dropped — And Open Source Is Now 3 Points From the Frontier

Kimi K3 is the successor to Kimi K2, the model that put Moonshot—a Beijing-based lab backed by Alibaba—on every founder's radar last year. K3 is a mixture-of-experts model with 2.8 trillion total parameters spread across 896 experts, native vision, and a 1-million-token context window. Two in-house architecture bets power it: Kimi Delta Attention, a hybrid linear attention mechanism that keeps million-token contexts affordable, and attention residuals, which Moonshot credits for consistent scaling gains.

The launch was timed a week ahead of the World AI Conference in Shanghai, and the model is already live on kimi.com, the mobile apps, and OpenRouter. The open-weights release is promised by July 27—the first open model in the 3-trillion-parameter class, if Moonshot delivers on the date.

The scoreboard is the story. K3 scores 57 on the Artificial Analysis Intelligence Index, behind only Claude Fable 5 (60) and GPT-5.6 Sol (59). On Moonshot's own benchmark suite, it beats last-generation flagships— Claude Opus 4.8 (max) and GPT-5.5 (high)—on most tests while losing to the newest frontier tier. In coding, it placed top-three across six benchmarks, led SWE Marathon and Program Bench outright, and trailed GPT-5.6 Sol on Terminal Bench 2.1 by half a point. Early hands-on reports back this up: developers on Hacker News called it "in the same ballpark as Fable or Sol," and Simon Willison noted a "notable improvement" on his infamous pelican-on-a-bicycle test.

Here's the part most launch coverage buries. K3's API costs $3 per million input tokens and $15 per million output tokens ($0.30 for cached input). That's Claude Sonnet-tier pricing, and roughly 4x its predecessor K2.6, which ran $0.95 in and $4 out. Simon Willison called it the most expensive model a Chinese lab has ever released. The Decoder's framing was blunter: the end of super cheap Chinese AI.

On real workloads the math is still competitive—Artificial Analysis measures K3 at about $0.94 per completed task, in the same band as GPT-5.6 Sol and roughly half of Claude Opus 4.8. But frontier-adjacent capability now costs frontier-adjacent money, whoever ships it:

Because the value of open weights was never really the discount. It's control.

When Anthropic pulled Fable 5 and Mythos 5 overnight earlier this year, every founder building on cloud AI got a free lesson in platform risk: a model you depend on can be deprecated, rate-limited, or re-priced without your consent. Weights on your own infrastructure can't be. Once K3's weights ship, US providers like Together, Fireworks, and Baseten will host them within days—solving the China-routing problem and creating hosting competition that Moonshot's own API has to price against.

And that's the second, bigger effect: open weights cap everyone's prices—including the models you actually use. Every time an open model lands within a few points of the frontier, closed labs have to justify their premium with reliability, efficiency, and trust rather than raw capability. The AI pricing war we covered in February was started by exactly this dynamic. You benefit from K3's existence even if you never send it a single token.

The gap that remains is real, though. The last three points are the hardest ones: hallucination control, reasoning efficiency, and the reliability that lets an agent run for an hour unsupervised. That's exactly where the frontier labs will dig their moat—and where K3, with its 51% hallucination rate on long-horizon work, isn't there yet.

The next two weeks tell you most of what you need to know. July 27 is the promised weights date—watch whether it lands on time and under what license, because an "open" release with a restrictive license is a different product. DeepSeek V4's general availability is expected within days, which will reset the budget end of the market again. And the frontier labs won't sit still: when an open model gets this close, the historical response is a price cut, a new tier, or both.

The bigger arc is clear. Eighteen months ago, open models trailed the frontier by a year or more. K3 lands three points off the lead while beating the flagships of six months ago. By the end of 2026, "open vs. closed" won't be a capability debate—it'll be a reliability, efficiency, and trust debate. Plan your stack for that world: the intelligence is becoming a commodity, and your moat is whatever you build on top of it.