Breaking

Catnip Unveils MaineCoon, a 22B Real-Time Generation Model

June 18, 2026 at 02:05 EDT

Media Generation
Foundation Models
Open Source

Catnip (@catnips_ai) on June 16, 2026 unveiled MaineCoon, a 22-billion-parameter real-time model that keeps a text-generated AI character on screen continuously without pausing. It generates audio and video simultaneously and can stream for unlimited duration, delivering what the company calls a "live AI character." Details

June 16, 2026 · Catnip

MaineCoon: An AI Character That Stays Alive As You Talk To It

A 22-billion-parameter real-time audio-visual model that turns text into an interactive live character — streaming unlimited synchronized video, audio and expression instead of a short fixed clip.

22B

parameters, real-time audio-visual model

<1s

to render the first frame

47.5

FPS peak on a single H100 GPU

1000s+

continuous streaming, and beyond

Throughput vs existing models

Internal tests: roughly 7× faster than existing audio-visual models.

1×

existing models

7×

MaineCoon

Generation cost: well below $0.001 per second

From watching → to talking to

Conventional video

Renders a short fixed clip, then stops

→

Turn-based chat

Waits, then replies in turns

→

MaineCoon

Continuous perception, a live character present with you

Early signal

State-of-the-art on the new SocialVideo Bench, surpassing seven open models in quality and speed. Released under Apache 2.0 — open weights, with developers calling it a natural, conversational presence.

Still to verify

At early-access stage. Extended real-world use, general-user reviews and added benchmarks await verification. No quality collapse or latency issues reported so far.

Continue reading

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