The takeaway: Moonshot AI's Kimi K3 is a genuinely frontier-class launch — 2.8 trillion parameters, independently scored fourth of 189 models tracked. It is also a case study in watching a launch page get edited in real time: "open-source" became "open," a weights promise went from "coming days" to a hard date of 27 July, and a headline benchmark number quietly changed. Until the weights actually land, the biggest "open" model in the world is — by the independent scorekeeper's own label — proprietary.

What actually launched

On Thursday afternoon, Moonshot AI released Kimi K3, the Beijing lab's new flagship. The verified spec sheet, from the company's own tech blog: a 2.8-trillion-parameter Mixture-of-Experts model activating 16 of 896 experts, built on Moonshot's Kimi Delta Attention and Attention Residuals architectures, with native vision and a 1-million-token context window. Moonshot claims the design yields "an approximate 2.5× improvement in overall scaling efficiency compared to Kimi K2." Active parameter count is not stated.

API pricing is $0.30 per million tokens for cache-hit input, $3.00 for cache-miss input, and $15.00 for output. The launch Hacker News thread passed 1,200 points overnight.

And the launch copy is unusually candid in one respect: Moonshot itself says K3 "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol."

The word "open" did some travelling

Here is the timeline, reconstructed from readers who kept the original page open in their browsers — every quote below is preserved verbatim in the launch thread.

The original launch copy said: "Kimi K3 is the first open-source model to reach the 2.8-trillion-parameter scale." It also said the full model weights "will be released in the coming days."

Mid-afternoon, the claim vanished. One reader: "They've removed the paragraph about releasing model weights." Another: "Any mention of 'open' has been struck from the literature for this model (it was present an hour ago)... At this pricing, I'll be surprised if it's open."

By evening, the language settled — and it is what the blog says now. K3 is "the first open model to reach 2.8 trillion parameters" — open model, not open-source — and "the full model weights will be released by July 27, 2026."

The difference between those framings is not pedantry. "Open-source" implies a licence; no licence has been named for K3, and its predecessors sit on Hugging Face under a custom "other" licence tag, not a standard open licence. "Open model" with a future weights date is a promise, not a property. As of this morning there is no K3 repository on Moonshot's Hugging Face page — the newest upload remains Kimi K2.7-Code.

One more thing moved during the edits: the launch copy quoted in the thread said K3 "scores 1687" on the GDPval-AA v2 leaderboard, ranking "behind only Claude Fable 5 Max and GPT-5.6 Sol Max." The benchmark table on the current page lists K3's GDPval-AA v2 Elo at 1668.0. The ranking claim survives either way — but a headline number changing between versions of a launch page, without annotation, is exactly why we cite archived quotes with IDs.

The independent numbers

Artificial Analysis has already scored K3: 57 on its Intelligence Index, #4 of 189 models tracked, with only three proprietary frontier models above it. Time-to-first-token measured 1.99 seconds, based on Kimi's API. And AA's classification is the crispest statement of where things stand: K3 is listed as a proprietary model — in the site's own words, "the model weights are not publicly available."

That classification is what makes 27 July matter. The highest-ranked model in AA's open-weights cohort today is GLM-5.2 (max), which scores 51 — #1 of the 97 open-weights models it is compared against. Thinking Machines' Inkling, which we covered when its independent numbers landed, scores 41. If K3's weights arrive as promised and AA flips the label, a 57 doesn't nudge the open-weights ceiling — it obliterates it, and the "leading open model" conversation restarts from a new baseline. If they don't arrive, a 2.8-trillion-parameter press release becomes its own story.

What the early hands say

Reaction in the launch thread runs hot and cold — sometimes in the same account. One developer's early hands-on update: "the subscription limits are pretty brutal... But the model itself is amazing. I think I might put this above Opus 4.8." The same reader later added that in a blind test they were "not sure I could tell the difference between this and Fable."

An independent third-party benchmarker was cooler: K3 landed "between GPT-5.6 Terra and GPT-5.6 Sol" on their suite — "~30% better than Kimi K2.6, but a lot slower and more expensive" — and they hit tool-calling schema rejections and timeouts along the way.

And the standing scepticism got its airing: "the benchmarks aren't going to tell the whole story," one commenter wrote, voicing the industry-wide suspicion that benchmark material leaks into training data. That concern is generic to every launch, not specific evidence against this one — but at a moment when the launch page's own numbers moved mid-day, it lands harder than usual.

What we're watching

Three dates and documents: 27 July, when the weights are due; the technical report Moonshot says will accompany them; and the licence text, which will decide whether "open" means open-source, open-weights-with-strings, or something else. Our field guide to what 'open' actually means in AI covers the taxonomy; today's Daily Update has the day's wider picture, in which a Chinese lab's promissory open crown landed the same week 29 countries signed an AI-governance organisation into existence in Shanghai.

The crown is promissory until the weights exist. Ten days on the clock.