Omead Pooladzandi

Co-Founder @PrismML PSGD

[Omead Pooladzandi @HessianFree]

your spotify cache is bigger than our largest AI model.

Bonsai: 1-bit weights. 1.7B to 8B params. 14x compression vs bf16. 8x faster on edge. 256 MB to 1.2GB. Based on Qwen 3.

we just came out of stealth. intelligence belongs at the edge and we're going to put it there.

Apache 2.0.

we compressed intelligence. more coming. @PrismML

[PrismML @PrismML]

Today, we are emerging from stealth and launching PrismML, an AI lab with Caltech origins that is centered on building the most concentrated form of intelligence.

At PrismML, we believe that the next major leaps in AI will be driven by order-of-magnitude improvements in intelligence density, not just sheer parameter count.

Our first proof point is the 1-bit Bonsai 8B, a 1-bit weight model that fits into 1.15 GBs of memory and delivers over 10x the intelligence density of its full-precision counterparts. It is 14x smaller, 8x faster, and 5x more energy efficient on edge hardware while remaining competitive with other models in its parameter-class.
We are open-sourcing the model under Apache 2.0 license, along with Bonsai 4B and 1.7B models.

When advanced models become small, fast, and efficient enough to run locally, the design space for AI changes immediately. We believe in a future of on-device agents, real-time robotics, offline intelligence and entirely new products that were previously impossible.

We are excited to share our vision with you and keep working in the future to push the frontier of intelligence to the edge.

[Larry Dial @classiclarryd]

Building momentum at Marin! Upgrading from Dense -> 129B parameter MoEs -> architecture improvements -> optimizer improvements gives our pretraining recipe an estimated 6x cumulative learning speedup, accounting for MFU. Includes community contributions. https://t.co/5dPB9uBiSp

[Omead Pooladzandi @HessianFree]

@norxornor @stochasticchasm @wen_kaiyue @dlwh Relu^2.3 is goated

[Omead Pooladzandi @HessianFree]

@varunneal @CevherLIONS the best way isnt clear yet. i usually just say well grads/momentum and hvps are going into a system that should learn to whiten both. then we just precond the momentum.

[Omead Pooladzandi @HessianFree]

@SeunghyunSEO7 @torchcompiled initializing zero is equivalent to initializing with orthogonal for muon

[Omead Pooladzandi @HessianFree]

@wen_kaiyue We’ve clearly over biased ourselves towards Adam in the wild and in the modded nano GPT speed run we’ve over biased ourselves towards muon.

[Omead Pooladzandi @HessianFree]

@KyleLiang5 congrats man!

[Kaizhao Liang @KyleLiang5]

it's a pretty good model 😮‍💨 and it's finally out

https://t.co/ypFJf5A5hw

[Unsloth AI @UnslothAI]

You can now train 120B+ parameter models locally on a laptop! 🔥

We collabed with NVIDIA and Microsoft to bring LLM training on the 128GB unified memory RTX Spark laptop!

[Omead Pooladzandi @HessianFree]

@MainzOnX Xilin Li does does the bulk of his work in complex128 for representing speech. That's why PSGD can optimize directly the complex domain.

[Michael Dell 🇺🇸 @MichaelDell]

We have the first @DellTech + @nvidia Vera Rubin NVL72 @CoreWeave. Here we go! 🚀

[Omead Pooladzandi @HessianFree]

novograd is still a good optimizer

[Omead Pooladzandi @HessianFree]

not bad at all

[Omead Pooladzandi @HessianFree]

@norxornor I bias more towards sgd at the start and then allow for more dynamic at the end. If I know I'm going to have a lot of noise I will also adjust eps.

[Omead Pooladzandi @HessianFree]

@norxornor I usually schedule eps

[Omead Pooladzandi @HessianFree]

@SemiAnalysis_ @eraznafre skill issue

[Omead Pooladzandi @HessianFree]

@agarwl_ How is this from 2024

[rohan anil @_arohan_]

Starting with Neolab tradition of writing our first blogpost. It’s a very good blogpost sir.

[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

[Omead Pooladzandi @HessianFree]

@danielhanchen @UnslothAI @marksaroufim @Microsoft Fun see you there!

[wesleimade @wesleimade]

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