Kimi K3 — крупнейшая открытая ИИ-модель уровня Opus 4.8 по цене Sonnet 5

In July 2026, the Chinese lab Moonshot AI released Kimi K3 — an open language model with 2.8 trillion parameters built on a Mixture of Experts architecture, of which about 50 billion are active per token. This is the largest open model ever published.

What is Kimi K3

Kimi K3 is Moonshot AI's flagship open model, built on a Mixture of Experts (MoE) scheme. The designation 2.8T-A50B breaks down as 2.8 trillion parameters in total and about 50 billion active at each generation step. This design gives the model the "knowledge" of a giant model, but in terms of compute cost per request it behaves closer to a dense model with ~50 billion parameters.

Key facts from the announcement:

• Name — Kimi K3, developer — Moonshot AI (creators of the Kimi lineup)
• Architecture — Mixture of Experts, 2.8 trillion parameters total, ~50 billion active (notation 2.8T-A50B)
• Status — the largest open model at the time of release
• Quality — claimed to be at the level of Claude Opus 4.8
• Price — comparable to Claude Sonnet 5
• Predecessor — the open model Kimi K2

Opus 4.8 level at Sonnet 5 price

Kimi K3 is positioned as a model in the Claude Opus 4.8 class in terms of quality, at a price on the level of Claude Sonnet 5 — a flagship result for the money of a mid-tier plan. It's precisely this "price/quality" ratio that was put in the headline of the breakdown by Latent Space in its AI News newsletter.

For the market, this is a direct challenge to closed flagships: if an open model with public weights genuinely reaches the level of top-tier Opus across a range of task classes, teams gain the option of deploying it on their own infrastructure — without paying per token and without sending data to an external API.

"Another great week for open models continues," reads the AI

News newsletter issue from Latent Space.

Why this matters for the market

The release of Kimi K3 continues a wave of strong open-weight releases. In recent cycles, the race for open models has been set by Chinese teams — Moonshot with its Kimi lineup, as well as DeepSeek and Qwen — and each new flagship has narrowed the gap with the closed systems of OpenAI, Anthropic, and Google.

A scale of 2.8 trillion parameters moves this race into a new weight class: public weights of this size simply hadn't been released before. An important caveat: the "Opus 4.8 class" claim, at launch, remains a positioning statement rather than a verified independent benchmark — real numbers will come from community measurements.

What this means

Open models have moved right up to the flagship level of closed labs — and they're doing it at the price of a mid-tier plan. For developers and companies, this expands the options: top-tier quality can now be obtained not only through the APIs of major vendors, but also on one's own servers.

Frequently Asked Questions

Who developed Kimi K3?

Kimi K3 was created by the Chinese AI lab Moonshot AI — the same team that released the previous open models in the Kimi lineup, including Kimi K2.

What does the notation 2.8T-A50B mean?

This describes the Mixture of Experts architecture: 2.8T is the model's total number of parameters (2.8 trillion), A50B is the number of active parameters per token (about 50 billion). At each step, only a portion of the "experts" is engaged, so the model computes faster than a dense network of the same size.