Moonshot выпустила Kimi K3 — крупнейшую открытую модель на 2,8 трлн параметров

Chinese lab Moonshot AI unveiled Kimi K3 on July 16, 2026 — a language model with 2.8 trillion parameters, which it calls the first "open model in the 3-trillion class." Open weights are promised to be released by July 27, 2026; for now, the model is only available through the company's website and API.

What is known about Kimi K3

Kimi K3 is Moonshot AI's largest model to date: 2.8 trillion parameters, more than twice the size of the previous trillion-parameter Kimi. In size it surpassed DeepSeek V4 Pro's 1.6 trillion parameters and, as Moonshot claims, became "the most capable model at this moment."

• 2.8 trillion parameters — the first "open model in the 3-trillion class," surpassing DeepSeek V4 Pro (1.6 trillion)
• Price: $3 per million input tokens and $15 per million output tokens — Claude Sonnet-level pricing
• The previous Kimi K2.6 cost $0.95 and $4 per million tokens — the price rose several times over
• Open weights promised by July 27, 2026
• Elo 1547 in the long-horizon knowledge work test, +732 points over Kimi K2.6

How strong is it?

According to Moonshot's own benchmarks, Kimi K3 beats Claude Opus 4.8 max and GPT-5.5 high but falls short of Claude Fable 5 and GPT-5.6 Sol. In an independent evaluation by Artificial Analysis, the model took second place in the long-horizon work test, coming in behind only Claude Fable 5.

"In our private long-horizon knowledge work evaluation,

Kimi K3 achieves an Elo of 1547 — 732 points above Kimi K2.6 and second only to Claude Fable 5," the Artificial Analysis report states.

According to Artificial Analysis data, the cost per task for Kimi K3 is $0.94 — roughly the same as GPT-5.6 Sol ($1.04) and half the cost of Claude Opus 4.8 ($1.80). At the same time, the model uses 21% fewer output tokens than Kimi K2.6. On the Arena.ai leaderboard for frontend coding, Kimi K3 took first place, even surpassing Claude Fable 5.

Why the price is record-breaking

Kimi K3 is the most expensive model released by a Chinese lab: $3 per million input tokens and $15 per million output tokens. That's the level of Anthropic's Claude Sonnet line and a sharp jump compared with the previous Kimi K2.6, which cost $0.95 and $4 per million tokens.

The price increase comes together with a size increase: 2.8 trillion parameters — more than twice the size of the previous generation's trillion-parameter Kimi. Moonshot is departing for the first time from the dumping strategy typical of Chinese models and is positioning K3 as a flagship rather than a cheap alternative to Western labs. All the published figures are Moonshot's self-reported benchmarks and a preliminary independent evaluation, so K3's final standing will become clear once the weights are released.

What this means

Chinese labs are no longer competing on price alone: Moonshot has released the largest open model and, for the first time, positioned it as a Western-flagship-level model. If the weights are indeed released on July 27, 2026, developers will get a 2.8-trillion-parameter model they can run on their own infrastructure.

Frequently asked questions

When will Kimi K3's open weights be released?

Moonshot AI promises an open release of the weights by July 27, 2026. Until then, the model is only available through the company's website and API.

How much does Kimi K3 cost via the API?

Kimi K3 costs $3 per million input tokens and $15 per million output tokens — that's Claude Sonnet-level pricing and the highest price among models from Chinese labs.

How does Kimi K3 differ from DeepSeek V4 Pro?

Kimi K3 is larger: 2.8 trillion parameters versus 1.6 trillion for DeepSeek V4 Pro. It is precisely because of this size that Moonshot calls K3 the first "open model in the 3-trillion class."