Is Kimi K3 released yet?

Not officially as of July 15, 2026. Moonshot has published a "K3 launch" promotion on the Kimi Open Platform dated July 15, but no model card, benchmarks, pricing, or weights have been posted. The specifications circulating come from press reports and insiders, not from Moonshot.

Will Kimi K3 be open-weight?

Probably, but Moonshot has not confirmed it for K3 specifically. Every Kimi model from K1 through K2.7 Code shipped as open-weight, so K3 following the same pattern is the reasonable expectation rather than a stated fact.

How much will Kimi K3 cost?

Moonshot has not announced pricing. A third-party projection puts it near its predecessor, roughly $0.80 to $1.20 per million input tokens and $3.00 to $4.50 output, based on Kimi K2.7 Code rates of $0.95 in and $4.00 out. Treat any K3 price as unconfirmed until Moonshot posts it.

Does Kimi K3 really have 2.5 trillion parameters?

That figure comes from press reports and unnamed insiders, not Moonshot. It would roughly double the largest open-weight models shipping today, which is possible but unverified. Wait for an official model card before treating the parameter count as real.

Should I switch to Kimi K3 now?

You cannot yet, because it is not released. If you need a cheap open-weight coding model today, use Kimi K2.7 Code, GLM-5.2, or DeepSeek V4, which are all shipping and benchmarked. Keep your integration model-agnostic so you can swap in K3 as a config change once it lands.

Kimi K3: What Is Confirmed and What Is Rumored

The screenshot going around is real. A Kimi Open Platform page is advertising a limited-time “K3 launch” promotion dated to begin July 15, 2026, which is why half of AI Twitter spent the weekend calling Kimi K3 imminent. What that page does not contain is a model card, a benchmark, or a price. So before you rewrite your agent config around it, here is the honest split: what Moonshot has actually confirmed, and what is still a rumor wearing a spec sheet.

Key takeaways

Confirmed: a Kimi Open Platform promotion points to a K3 launch around July 15, 2026. That is a launch signal, not a spec sheet.
Not confirmed: the headline numbers (2.5 trillion or more parameters, 1M-token context, native multimodality) trace to press reports and insiders, not a Moonshot announcement.
Likely but unstated: open weights. Every Kimi model from K1 through K2.7 shipped open-weight, so K3 probably will too, but Moonshot has not said so for K3.
What matters for you: K3 extends a cost gap that already exists. Its predecessor Kimi K2.7 Code runs at $0.95 in and $4.00 out per million tokens, a fraction of a closed frontier model.
The move right now: do not migrate on a rumor. Wait for the weights and a real benchmark. This page updates to a full review the moment those land.

Is Kimi K3 out yet?

As of July 15, 2026, Kimi K3 is not officially released. Moonshot has published a launch promotion on the Kimi Open Platform dated to July 15, but no model card, benchmark results, pricing, or weights have been posted. Every specification circulating (parameter count, context window, multimodality) comes from press reports and unnamed insiders, as covered by outlets including AIbase, not from Moonshot itself. Treat the launch as happening and the specs as unconfirmed.

That distinction matters more than usual here, because the numbers being repeated are big enough to move buying decisions, and being wrong about them is expensive.

Kimi K3: confirmed vs rumored

Detail	Status	Source
K3 launch promotion dated July 15, 2026	Confirmed (promo page)	Kimi Open Platform, via press
Open weights	Expected, not stated for K3	Pattern from K1–K2.7
~2.5 trillion or more parameters	Rumored	Press / insiders
1M-token context window	Rumored	Press / insiders
Native multimodality	Rumored	Press / insiders
Pricing	Not announced	Third-party projection only

The parameter figure is the loudest and the softest. Reports put K3 at 2.5 trillion parameters or more, framed as surpassing domestic rivals, which would roughly double the largest open-weight models shipping today. It might be true. It is also exactly the kind of round, dramatic number that leaks before a launch and quietly shrinks by release day. Until Moonshot posts a model card, it is a rumor.

What K3 is built on: the K2.7 Code lineage

The reason anyone cares about K3 is not the rumored parameter count. It is the track record of the model right below it.

Kimi K2.7 Code, Moonshot’s current coding model, shipped June 12, 2026 with stronger long-horizon coding and about 30% lower thinking-token use than its predecessor. Its whole design goal is finishing a real software task across a long run of tool calls without losing the thread. On Artificial Analysis’s general Intelligence Index it scores 42, which looks unremarkable until you remember that index is a broad composite: a model tuned narrowly for agentic coding gets undersold by physics and trivia questions. The thing to watch with K3 is not whether it wins a general benchmark. It is whether it finishes the task.

Moonshot has also kept every Kimi release open-weight, which is the pattern that makes the open-weight economics work in the first place. For where K2.7 Code and its rivals sit in the current field, see the best open-weight AI models in 2026, and for the picks by job, the AI coding agents comparison.

The number that actually moves: cost

Strip away the parameter rumor and here is what K3 is really about. The Kimi line is cheap, and cheap that stays good is what has pulled real workloads onto Chinese open-weight models this year.

Open-weight coding models cost a fraction of closed frontier models
Item	Value
DeepSeek V4 Pro	$0.87
Kimi K2.7 Code	$4
GLM-5.2	$4.4
Claude Opus 4.8	$25
GPT-5.5	$30

Kimi K2.7 Code lists at $0.95 per million input tokens and $4.00 per million output tokens. Claude Opus 4.8 runs $25 on output; GPT-5.5, $30. That is the gap teams noticed, and it is why K2.6 racked up hundreds of billions of processed tokens on OpenRouter before K3 was even a rumor. A third-party projection pegs K3 in a similar band to K2.7, around $0.80 to $1.20 input and $3.00 to $4.50 output, but that is a projection, not a Moonshot price.

The per-token price is also not the whole story, which is the part most coverage skips. A coding agent burns tokens by the million per finished task, so caching and tool-call reliability move the real bill as much as the sticker rate. If K3 keeps the K2 line’s aggressive cache discounts, its effective cost per task drops well below what the rate card suggests. We walk through why in why Chinese AI models are so cheap, and you can model a specific model against your own workload in the AI models tracker.

Should you wait for K3 or ship on K2.7 now?

Ship on what exists. If you need a cheap, capable open-weight coding model today, Kimi K2.7 Code is real, benchmarked, and priced, and so are its rivals GLM-5.2 and DeepSeek V4. If you specifically want Moonshot and can wait, keep your integration model-agnostic so swapping K2.7 for K3 is a config change, not a rebuild. What you should not do is delay a decision or rearchitect around specs that Moonshot has not confirmed. For a model-picker framing across the current field, the best models for the Hermes agent breakdown does the sorting by job.

Frequently asked questions

Is Kimi K3 released yet?: Not officially as of July 15, 2026. Moonshot has published a "K3 launch" promotion on the Kimi Open Platform dated July 15, but no model card, benchmarks, pricing, or weights have been posted. The specifications circulating come from press reports and insiders, not from Moonshot.
Will Kimi K3 be open-weight?: Probably, but Moonshot has not confirmed it for K3 specifically. Every Kimi model from K1 through K2.7 Code shipped as open-weight, so K3 following the same pattern is the reasonable expectation rather than a stated fact.
How much will Kimi K3 cost?: Moonshot has not announced pricing. A third-party projection puts it near its predecessor, roughly $0.80 to $1.20 per million input tokens and $3.00 to $4.50 output, based on Kimi K2.7 Code rates of $0.95 in and $4.00 out. Treat any K3 price as unconfirmed until Moonshot posts it.
Does Kimi K3 really have 2.5 trillion parameters?: That figure comes from press reports and unnamed insiders, not Moonshot. It would roughly double the largest open-weight models shipping today, which is possible but unverified. Wait for an official model card before treating the parameter count as real.
Should I switch to Kimi K3 now?: You cannot yet, because it is not released. If you need a cheap open-weight coding model today, use Kimi K2.7 Code, GLM-5.2, or DeepSeek V4, which are all shipping and benchmarked. Keep your integration model-agnostic so you can swap in K3 as a config change once it lands.

Sources

AIbase (2026). China’s Large Model Scene: Kimi K3 to Be Released This Month with Parameters Reaching 2.5 Trillion (press report; unconfirmed by Moonshot). news.aibase.com
KuCoin (2026). Kimi K3 to launch this month with 2.5 trillion parameters, surpassing domestic rivals (news flash; press report). kucoin.com
Artificial Analysis (2026). Kimi K2.7 Code: Intelligence, Performance and Price Analysis (independent benchmark and pricing). artificialanalysis.ai
TokenMix (2026). Kimi K3 Developer Integration Guide (secondary; third-party pricing projection, not a Moonshot price). tokenmix.ai