Nº 53AI ModelsJuly 24, 202610 min read

Kimi K3: The Largest Open Model Ever Does Not Exist Yet. It Is Due Monday.

Moonshot AI's 2.8 trillion parameter Kimi K3 beat Claude at frontend code on an independent leaderboard, and the press already calls it the largest open-weight model in history. One catch: the weights are not out until July 27. Here is what is verified, what is vendor slideware, and what a 28-trillion-parameter typo tells you about AI journalism.

By T.W. Ghost

The headlines say Moonshot AI just released the largest open-weight model in history. The headlines are early.

Kimi K3, launched July 16 with no keynote and no fanfare, is a 2.8 trillion parameter model that Moonshot calls "the world's first open-source model in the 3-trillion-parameter class." As of this writing, that sentence is a promise, not a fact: Artificial Analysis still classifies K3 as proprietary, Moonshot's Hugging Face organization still tops out at Kimi K2.7, and the company's own documentation says the full weights "will be released by July 27, 2026."

That is Monday. Until then, the largest open model ever is the largest *announced* open model ever, and the difference matters. Here is the verified record.

Who Is Moonshot, and Why This Is Their Shot

Moonshot AI is the Beijing lab behind Kimi, one of China's most popular consumer AI assistants. Their Kimi K2 series earned real respect among developers as capable open-weight workhorses, but always a tier below the frontier. K3 is the company's claim to the top table: not just China's biggest model, but the biggest anyone has ever offered to hand over.

The architecture, per Moonshot's documentation: 2.8 trillion total parameters in a sparse mixture-of-experts design that activates 16 of 896 experts per token, a hybrid attention scheme the company calls Kimi Delta Attention, and a vendor-claimed 2.5x scaling efficiency over K2. A 1,048,576-token context window, four times K2's, with native vision input, always-on reasoning, tool calling, and strict JSON output.

All of those numbers are Moonshot's own until Monday. That is the whole point of open weights: on July 27, anyone can count the parameters themselves.

The 28 Trillion Parameter Typo, an Autopsy

While researching this piece we kept meeting a claim that K3 has 28 trillion parameters, ten times the real figure. We traced it: the earliest "28 trillion" appears in a crypto-news flash headline that dropped the decimal from 2.8, and the error then propagated into other outlets' URL slugs, some of which still say "28-trillion" while their own article text says 2.8.

A related figure, "A50B" (50 billion active parameters), appears in community shorthand everywhere and in Moonshot's documentation nowhere. It is an estimate derived from the 16-of-896 expert ratio. Possibly right, never published.

We mention this not to dunk on anyone but because it is this site's whole thesis in miniature: launch-week numbers mutate as they travel. Read the primary source.

The Mirror: Meta Locks the Door, Moonshot Opens One

The timing writes its own story. In April, Meta killed the Llama line and went closed. Two weeks ago it opened a paid API instead, renting access to a model whose weights you will never see, US developers only.

The same month, a Beijing lab is preparing to hand the world the largest set of open weights ever released, under an announced modified-MIT license. The open-weights flag Meta dropped is being picked up on the other side of the Pacific, at 2.8 trillion parameters. Whatever else is true about K3, the symbolism is not subtle, and the license text (still unpublished as of this writing) is the detail to check on Monday.

The Benchmark That Survived Verification

One K3 result is independently confirmed, and it is a good one: #1 on Arena AI's Frontend Code Leaderboard at 1,679 Elo, ahead of Claude Fable 5 at 1,631. That is a community pairwise-voting leaderboard, not vendor self-reporting, and it represents a 17-place jump from Kimi K2.6's #18. For frontend generation specifically, real votes by real users currently prefer K3's output to the best model on the planet's.

Now the other side of the ledger. The broader narrative you may have seen, that K3 ranks just behind Claude Fable 5 overall and beats Claude Opus 4.8 and GPT-5.5 across the board, failed our verification. Those claims trace back to Moonshot's own launch charts, and the one independent composite score being circulated did not check out. Outside the Frontend Code Arena result, treat every K3 superiority claim as vendor-reported until the weights let independents run their own evals, which is days away.

Simon Willison, reviewing the launch, flagged the same epistemic gap: at release there was no public model card, no license file, and no downloadable weights, so nobody outside Moonshot could verify much of anything.

The Price of Thinking

K3's pricing, verified on Moonshot's own pricing page:

	Kimi K3 API
Input (cache hit)	$0.30 per 1M tokens
Input (cache miss)	$3.00 per 1M tokens
Output	$15.00 per 1M tokens
Context	Flat pricing across the full 1M window

Two things are true at once. This is the most expensive model a Chinese lab has ever shipped, towering over DeepSeek V4 Pro ($0.435/$0.87) and GLM 5.2 ($1.40/$4.40). And it is still roughly half the per-task cost of Western frontier models: Willison measured about $0.94 per task against roughly $1.80 for Claude Opus 4.8.

The catch lives in that word "task." K3's reasoning is always on, and reasoning tokens bill as output at $15 per million. Willison's now-famous example: asking for a simple SVG drawing consumed 95 input tokens and 16,658 output tokens, of which 13,241 were the model thinking, about $0.25 for one image. Cheap per token and cheap per task are different claims; benchmark your own workload before believing either. One more hedge: Moonshot's docs describe multiple reasoning-effort levels, but at launch only the maximum appeared to be exposed. Expect this to be in flux.

Using It Today, and What Monday Actually Changes

What you can do right now:

Moonshot API: OpenAI-compatible, model id kimi-k3, unlocked after a $1 minimum top-up.
OpenRouter: listed at $3/$15, reportedly without prompt caching, so you pay sticker price.
The Kimi app: new consumer subscriptions were paused in mid-July amid a GPU capacity crunch, which is also the main schedule risk hanging over the July 27 date.

And a reality check on what the weights drop means for you personally: nothing about running it at home. A 2.8T mixture-of-experts model is a datacenter problem; no desktop, and no reasonable homelab, will host it. What open weights actually buy is bigger than self-hosting: independent labs can finally verify the benchmarks, distillers can shrink it into models you *can* run, enterprises get sovereignty options, and every closed lab's pricing meeting gets a little more uncomfortable.

What to watch Monday: the weights landing (or not) on Moonshot's Hugging Face organization, the actual license text against the announced modified MIT, and the first independent benchmark runs.

Who Should Use Kimi K3

Use it today if you are building agentic or coding workloads, especially frontend generation where its one verified benchmark wins outright, and you want near-frontier quality with a 1M context at half of Western frontier task costs. Budget for the reasoning-token burn.
Wait for Monday if you care about verified capability claims. The weights release converts vendor numbers into checkable ones, and this site will update accordingly.
Skip it if you need a consumer app right now (subscriptions paused) or if your workload is latency-sensitive chat where always-on reasoning fights you.
For the record: Claude Fable 5 remains the verified overall leader, Gemini 3.6 Flash the efficiency play, and Muse Spark 1.1 the closed newcomer K3 is implicitly aimed at.

The honest summary: Kimi K3 has one independently confirmed crown, a genuinely historic open-weights promise with a Monday deadline, the highest price tag in Chinese AI history, and a set of launch claims that still need the verification only those weights can enable. We will be watching the Hugging Face page like everyone else.