Breaking

Moonshot AI Unveils Kimi K3, a 2.8-Trillion-Parameter Multimodal Model With a 1 Million-Token Context Window

July 16, 2026 at 15:06 EDT

Foundation Models
Open Source
Software Dev & Coding

Moonshot AI on July 16, 2026 released Kimi K3, a 2.8-trillion-parameter multimodal model with a 1 million-token context window that the company bills as the world's first open model in the 3-trillion-parameter class. The model went live immediately across the company's web and mobile apps, its coding tools, and an OpenAI-compatible API, with open weights slated to follow on July 27, 2026.

July 16, 2026 · Moonshot AI

Kimi K3: the first open model in the 3-trillion class

A 2.8-trillion-parameter multimodal model with a 1 million-token context window, live across web, apps and an OpenAI-compatible API — with open weights slated for July 27.

2.8T

Total parameters — largest open-weight model to date

Token context window (1,048,576) with 1M-token completions

6.3×

Faster decoding at million-token contexts (KDA attention)

API pricing — per million tokens

Positioned to undercut Western frontier models

Input

$15

Output

Automatic caching can drop cached-input pricing sharply.

Kimi Delta Attention (KDA)

Hybrid linear-attention mechanism — up to 6.3× faster decoding in million-token contexts.

Attention Residuals (AttnRes)

Roughly 25% higher training efficiency at less than 2% added cost.

Built for

Long-horizon agentic coding Long-document knowledge work Self-evolving workflows Vision-in-the-loop: images & video → code

Early praise

Game developers call its speed, quality and consistency impressive; the local-model community reacted with astonishment at a 2.8T open-weight release.

The caveats

Running it locally demands hardware well beyond consumer reach. Independent benchmarks are pending, and active parameters per token remain undisclosed.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.