PrismML

Centering AI research on efficiency.

[PrismML @PrismML]

Today we’re releasing 1-bit and Ternary Bonsai Image 4B.

A new family of image-generation models designed to run high-quality diffusion inference on local hardware: from laptops to phones.

[Sahin Lale @SahinLale]

After @tcarambat’s video on our model — s/o for the support of local AI — we noticed the outputs looked much worse than expected on Macs.

Turns out it was a text encoder padding bug. Fixed now, and the improvement is huge! See the before/after results.

Please update your Bonsai Image pipeline to get the best outputs from your prompts!

[Evan Walters @evaninwords]

Bug fix! Bonsai Image generations on local MacBook MLX will be even better quality.

Turns out how you pad text matters 😆 try it out! https://t.co/vJFTG18oNf

[AJ @ItsmeAjayKV]

First tests on my 12GB RTX 3060 with the new @PrismML Ternary Bonsai image 4B model are actually insane.

1408×704 in ~2.8s (4 steps),
peak VRAM under 7GB
extremely consistent

This might be the best text2image model you can run on small consumer cards rn.

Better benchmark is being done, will be sharing results soon.

[Rohan Mukherjee @roerohan]

Local models are THIS GOOD already?

[Ostap Kolinets @yklnss]

A huge deal for the future of local AI

[Harshil @harshil1712]

Love everything PrismML's been doing for small local models!
You can generate images, on your tiny computer or browser 🔥

[Daniel Yurkin @danyurkin]

running this locally on an iPhone is kinda crazy ngl!

big deal @PrismML

[Adrien Grondin @adrgrondin]

On-device image generation on iPhone that is actually fast and also high quality!

Impressive model by @PrismML.

[wesleimade @wesleimade]

Continuing my tests with Ternary Bonsai Image 4B. Same prompt with different seeds. Amazing work @PrismML!

[Crown 👑 @barackomaba]

This is a huge leap forward for high quality diffusion inference on local hardware. I've had to spend 40GB+ to get similar quality in the past. The software optimizations happening are just as important as the hardware ones, if not more important. The future is bright!

[Vibe Coding with Gene @with_gene2626]

The age of Ternary Models has truly begun...

[PrismML @PrismML]

RT @neuroblossom: also try their langauge models

[Ivan Fioravanti ᯅ @ivanfioravanti]

@PrismML ~10 seconds for 1024x1024 using the Ternary model on M5 Max. WOW!

[Shikhar @shikhr_]

1-bit diffusion models are a thing now. It can run locally on your phone and create images within seconds.

At this pace, in a year or so, we'll get local video generation on consumer hardware.

[wesleimade @wesleimade]

Testing the new bonsai model from @PrismML . I'm very impressed with this! Running 100% locally in my browser. https://t.co/QW7uvU1VBL

[David Brown @DaBrown95]

@PrismML Really impressive, ~60 seconds for 1024x1024 using the Ternary model on my 17 Pro Max!

[Kanu Gulati @Khosla Ventures @KanuGulati]

Try the Bonsai Studio app on your iPhone: https://t.co/aFSdTD43iB

@PrismML 👏👏

[Kanu Gulati @Khosla Ventures @KanuGulati]

Diffusion models can be compressed too, quite aggressively, without losing the magic.

Intelligence Density is the metric: more capability per bit, per watt, per device.

At 1.21GB, PrismML’s ternary model keeps 95% of FLUX.2 Klein 4B accuracy while shrinking footprint 6.4x.

That means generation can move directly into the product experience on local devices: cheaper, faster, and private by default!

Congratulations to the @PrismML team.
Proud investor moment 👊
cc @khoslaventures

[Xenova @xenovacom]

WTF?! This changes image generation forever! 🤯 PrismML just released Binary and Ternary Bonsai Image 4B!

That's right, 1-bit diffusion models are here. Only ~3GB in size (FLUX.2 Klein 4B is 16GB).

The most shocking part? It can run 100% locally in your browser. Try it now! 👇

[Erfanzar @eraznafre]

1-Bit LLM > 1-Bit diffusion > 1-Bit ...?