Breaking

Moonshot AI Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model That Edited Its Own Launch Teaser

July 17, 2026 at 01:33 EDT

Foundation Models
Open Source
Media Generation

Moonshot AI on July 16, 2026 released Kimi K3, a 2.8-trillion-parameter open-weight multimodal reasoning model, and drew immediate attention with a launch teaser the model reportedly edited itself from 56 source clips—work that would normally take an experienced human editor one to two days.

July 16, 2026 · Moonshot AI

Kimi K3: A 2.8-Trillion-Parameter Open Model That Edited Its Own Teaser

Moonshot's new open-weight multimodal reasoning model assembled its launch trailer from 56 raw clips autonomously — work a human editor would take one to two days to finish.

2.8T

total parameters (MoE)

source clips auto-edited into one teaser

1.05M

token context window

16 / 896

experts active per token

Coding benchmark: Kimi K3 vs Claude Fable 5

One community-run coding evaluation — higher is better

1,679

Kimi K3

1,631

Claude Fable 5

A narrow edge over a rival closed model — though on broader general-intelligence measures K3 still trails top-tier systems.

Pricing — per million tokens

$3.00

input · cache miss

$0.30

input · cache hit

$15.00

output tokens

A structure widely viewed as aggressive relative to comparable frontier systems.

What developers praise

Autonomous video editing with vision in the loop
3D game generation via Three.js, playable assets
Research slides and GPU kernel optimization
A milestone for open-weight models from China

The caveats

Some features (e.g. web search) still being updated
Full model weights not yet public — due July 27
Trails Claude Fable 5 & GPT-5.6 Sol on general intelligence

The real test ahead

If the weights ship on schedule, the bigger question is how Kimi K3 performs once it's downloadable, fine-tunable and deployable outside Moonshot's own infrastructure.

Continue reading

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