Kimi K3 Is the Largest Open AI Model Ever — With a Catch

Moonshot AI released Kimi K3 on July 16, 2026, and on paper it is the largest model ever pitched as "open": a 2.8-trillion-parameter mixture-of-experts system with a million-token context window that lands just behind Claude Fable 5 and GPT-5.6 Sol on independent leaderboards. The catch is that "open" is a promise, not a fact yet. As of July 17 there is no downloadable checkpoint, no license, and no model card. For now you can only reach Kimi K3 through Moonshot's website and paid API, with the full weights scheduled for July 27. Here is what actually shipped, how it performs, and whether it belongs in your stack this week.

What is Kimi K3?

Kimi K3 is Moonshot AI's new flagship large language model, built as a sparse mixture-of-experts network with 2.8 trillion total parameters that routes each token through just 16 of 896 experts, so only a small slice of the network runs per token. It ships with a one-million-token context window and a native multimodal design that handles text, images, and video in a single architecture. Moonshot aims it squarely at long-horizon coding, knowledge work, and agentic tool use rather than casual chat. The headline number — 2.8 trillion parameters — makes it the first model of its scale to be announced as open-weight, roughly a "3T-class" system in Moonshot's own framing. It is available now through the Kimi web app and API; the downloadable weights are the part still in transit.

The catch: "open" but you can't download it yet

This is the part the launch-day headlines skipped. Moonshot marketed K3 as an open model, but on release day there was no K3 checkpoint in its public repository, no license, and no model card. The company set an open-weight release date of July 27, more than a week after the announcement, which strongly suggests the weights simply were not ready when the marketing was. Until they land, "open" is a roadmap item. You cannot self-host it, fine-tune it, or audit it — you rent it through Moonshot's API like any closed model. That matters because the entire appeal of an open frontier model is control: running it on your own hardware, keeping data in-house, and avoiding per-token metering. None of that is available yet. If Moonshot ships permissive weights on schedule, K3 becomes genuinely significant. If the date slips or the license is restrictive, it is a closed model wearing an open badge. Judge it on July 27, not on the press release.

How Kimi K3 performs against GPT-5.6 and Claude Fable 5

On independent testing, K3 is excellent but not the new king. Moonshot's own reported figures are strong — Program Bench 77.8, GPQA-Diamond 93.5, MMMU-Pro 81.6, and BrowseComp 91.2 — but vendor benchmarks always flatter the vendor. The more useful read is Artificial Analysis, which places K3 just behind Claude Fable 5 and GPT-5.6 Sol on its composite intelligence index, at an Elo of roughly 1547. Notably, it reaches that level while using about 21% fewer output tokens than Moonshot's previous K2.6, which is a real efficiency gain for anyone paying per token. The short version: K3 is the strongest open-lineage model yet and a credible fourth behind the top US frontier systems, but it does not dethrone them. For agentic coding it is competitive; for the absolute ceiling on hard reasoning, Fable 5 and GPT-5.6 Sol still lead. That gap has been narrowing all year, and open models catching the frontier is a trend we covered when the last wave landed.

Model	Architecture	Context	Input $/1M	Output $/1M	Open weights
Kimi K3	2.8T MoE (16/896 experts)	1M	$3.00 ($0.30 cached)	$15.00	Promised July 27
GPT-5.6 Sol	Closed flagship	1M	$5.00	$30.00	No
Kimi K2.6	MoE (prior gen)	256K	$0.95	$4.00	Yes

What Kimi Delta Attention actually changes

The interesting engineering in K3 is not the parameter count, it is how Moonshot made a million-token context affordable to serve. The model introduces Kimi Delta Attention (KDA), a hybrid linear-attention scheme that Moonshot says delivers up to 6.3x faster decoding in million-token contexts than standard attention. A second technique, Attention Residuals, is claimed to raise training efficiency by roughly 25% at under 2% extra compute cost. Both matter more than a leaderboard point: long-context inference has been the expensive, slow part of running big models, and a 6x decode speedup is the difference between a usable million-token agent and one that stalls. If those numbers hold up under independent replication once the weights ship, KDA is the piece other labs will study. For now they are vendor claims, credible but unverified outside Moonshot.

What Kimi K3 costs to use

K3 is priced at $3 per million uncached input tokens, $0.30 per million cache-hit input tokens, and $15 per million output tokens through Moonshot's API. That is a notable jump from K2.6's roughly $0.95 input and $4 output, and it puts K3 in the same rough band as Anthropic's mid-tier Sonnet pricing rather than the bargain tier Moonshot used to occupy. It still undercuts GPT-5.6 Sol's $5 input and $30 output by a wide margin, so on a pure cost-per-quality basis K3 is aggressive against the top closed models — just no longer the cheap option among Chinese labs. The cache-hit rate of $0.30 is the number to design around: if your workload reuses long system prompts or documents, aggressive prompt caching turns K3 from mid-priced into genuinely cheap. For high-volume, repetitive agent loops that reuse context, the economics are strong. For one-off long-output generation, the $15 output rate adds up fast.

Why a Chinese open model rivaling US labs matters

Kimi K3 is the clearest sign yet that the open-versus-closed and US-versus-China gaps are closing at the same time. A Chinese lab shipping a 3T-class model that ranks fourth globally, and promising to release the weights, pressures the closed US frontier on both price and openness. If the weights arrive under a usable license on July 27, any company with enough GPUs can run a near-frontier model with no vendor in the loop — the same logic that makes teams pick self-hosted infrastructure over metered SaaS. That does not mean it is the right pick for everyone: near-frontier still trails frontier, and running a 2.8T model yourself needs serious hardware. But it resets the default question from "which closed API do we trust" to "do we even need a closed API." For teams already weighing the best AI coding tools on real work, K3 is now part of that conversation in a way no open model was a year ago.

Should you use Kimi K3 right now?

• Building agentic coding or long-context tools and want frontier-adjacent quality cheaper than GPT-5.6 Sol? K3's API is worth testing today, especially if you can lean on prompt caching.
• Waiting to self-host an open frontier model? Wait for July 27 and check the actual license before you commit — right now there is nothing to download.
• Need the absolute best reasoning? Stay on Claude Fable 5 or GPT-5.6 Sol; K3 is close but a step behind.
• On a tight budget doing simple work? K3 is no longer the cheapest Chinese model — its own K2.6 is far less per token.

For most teams the honest move is to try the API this week, and hold any self-hosting plans until the weights and license are actually public.

Frequently asked questions

Is Kimi K3 Chinese? Yes. Kimi K3 is built by Moonshot AI, a Beijing-based startup, and Kimi is its consumer and developer AI brand. The model rivals top US systems on public leaderboards, which is a large part of why its release drew attention.

Will Kimi K3 be open source? Moonshot has announced an open-weight release scheduled for July 27, 2026, but as of July 17 no weights, license, or model card had been published. Until those land, K3 is only usable through Moonshot's API and web app, and the exact license terms are unknown. Treat "open" as a promise to verify, not a shipped fact.

How much does Kimi K3 cost? Through Moonshot's API, K3 costs $3 per million uncached input tokens, $0.30 per million cache-hit input tokens, and $15 per million output tokens. That is more expensive than the previous Kimi K2.6 (about $0.95 input and $4 output) but well under GPT-5.6 Sol's $5 input and $30 output.

How is Kimi K3 different from Kimi K2? K3 is a larger 2.8-trillion-parameter mixture-of-experts model with a one-million-token context window, native multimodal input, and the new Kimi Delta Attention mechanism for faster long-context decoding. It scores higher on reasoning and coding benchmarks than K2.6 while using fewer output tokens, but it also costs more per token to run.

Sources

• MarkTechPost — Moonshot AI Releases Kimi K3: architecture (2.8T MoE, 16/896 experts, 1M context), Kimi Delta Attention 6.3x decode claim, Attention Residuals, and reported benchmark scores.
• Simon Willison — Kimi K3, and what we can still learn from the pelican benchmark: independent analysis of the open-weight status (no checkpoint/license/model card at launch, July 27 date), Artificial Analysis Elo ~1547, and API pricing versus K2.6.
• Kimi (Moonshot AI) — official site: product positioning for agentic coding and knowledge work, and API access.