Breaking

Moonshot AI Unveils Kimi K3, a 2.8-Trillion-Parameter Model Set to Open-Source Its Weights

July 19, 2026 at 00:45 EDT

Foundation Models
Open Source
AI Agents

Beijing-based Moonshot AI has released Kimi K3, a 2.8-trillion-parameter natively multimodal model that the company plans to publish as open weights by late July, positioning it as the first frontier system at the three-trillion-parameter scale to be made openly available. The model launched around July 16, 2026, and is currently accessible for free through the company's official Kimi interface, where users can select K3 from a dropdown and begin building immediately.

July 16, 2026 · Moonshot AI

Kimi K3: The First 3-Trillion-Parameter Frontier Model Going Open

Beijing's Moonshot AI unveils a 2.8-trillion-parameter, natively multimodal system built for long-horizon agentic work — with full open weights promised by late July, a move the financial press is calling a second "DeepSeek shock."

2.8T

total parameters — frontier scale

token context window

16 / 896

experts active per inference (LatentMoE)

Scaling Efficiency

2.5× more efficient than the K2 series

1×

2.5×

Stable LatentMoE, Kimi Delta Attention and Attention Residuals combine to push more capability per unit of compute.

API Pricing (per million tokens)

input

$15

output

$0.30

cached

Undercutting many comparable frontier offerings — free now via the Kimi interface, also on OpenRouter and Cloudflare AI.

Early Praise

Developers say K3 matches or beats leading closed models on practical coding and reasoning, with strong agentic software-engineering results and even playable game-generation demos — at a fraction of the cost.

Open Questions

Some caution that benchmark scores may reflect overfitting, and question how reliably K3 holds up across extended, multi-step agentic workflows — to be clarified by independent testing.

The Milestone — July 27, 2026

Full weights to be published as open source

This would make K3 the first three-trillion-parameter-class system released openly — a direct challenge to closed rivals from Anthropic and OpenAI, and a potential accelerant for downstream research well beyond Moonshot's own infrastructure.

Continue reading

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