Today we’re introducing Gemma 4 12B — our latest open model that brings advanced agentic reasoning, vision and audio directly to your laptop.
It delivers performance nearing our larger Gemma models with a much smaller total memory footprint, while being small enough to run locally with just 16GB of VRAM. It’s open and accessible for everyone to use under a permissive Apache 2.0 license.
This is all made possible by our new, unified architecture that removes separate multimodal encoders. Here’s how we did it 🧵
@econoar The answer to decentralization isn't Either/ Or. It's also not Now/ Later.
It's a user selected spectrum that exists NOW. I.e.
https://t.co/M1AoRmAGjE
@aoTheComputer
Very excited about this launch. Introducing:
➡️ PermawebOS nodes: A bootable, multi-architecture HyperBEAM instance.
➡️ LapEEs: Pretty Good Security via swarms of widely available commodity hardware.
➡️ Decentralized bundlers: A first alpha deployment for LapEEs.
Demanding the most recent high-end server chips or ~$100k as an entry ticket to contribute hardware just isn't a path to decentralized networks.
To actually decentralize the web we are going to need to enable wide participation. The PermawebOS is how we think we get there. Big day 🙂.
To be clear: This is a huge ship and it is definitively, categorically an alpha. We have been running and using LapEEs ourselves at @fwdresearch for weeks, but you shouldn't jump on this today unless you are ready for: A. Nothing to work as you expect. B. The very real possibility that you brick your machine.
Stability and ease-of-use will improve over time, so unless you love self-guided firmware flashing adventures it is best to hang tight for now.
Further down the track, it also looks like Android deployments of the PermawebOS may be possible, too, using kernel software isolation and some of their security hardware. Running a bundler on your phone that you genuinely do not know how to tamper with even if you wanted to is pretty remarkably.
Brick by brick.
Very excited about this launch. Introducing:
➡️ PermawebOS nodes: A bootable, multi-architecture HyperBEAM instance.
➡️ LapEEs: Pretty Good Security via swarms of widely available commodity hardware.
➡️ Decentralized bundlers: A first alpha deployment for LapEEs.
Demanding the most recent high-end server chips or ~$100k as an entry ticket to contribute hardware just isn't a path to decentralized networks.
To actually decentralize the web we are going to need to enable wide participation. The PermawebOS is how we think we get there. Big day 🙂.
To be clear: This is a huge ship and it is definitively, categorically an alpha. We have been running and using LapEEs ourselves at @fwdresearch for weeks, but you shouldn't jump on this today unless you are ready for: A. Nothing to work as you expect. B. The very real possibility that you brick your machine.
Stability and ease-of-use will improve over time, so unless you love self-guided firmware flashing adventures it is best to hang tight for now.
Further down the track, it also looks like Android deployments of the PermawebOS may be possible, too, using kernel software isolation and some of their security hardware. Running a bundler on your phone that you genuinely do not know how to tamper with even if you wanted to is pretty remarkably.
Brick by brick.
It's difficult to find many defensible ideas though.
Any "agent first" / "headless" ideas that are meaningfully big will just get eaten up by the labs
Any smaller, niche ideas are all fine but instnalty you now have 10+ competitors in weeks, whereas in 2010 you'd get a few years headstart
@garrytan The hiring tell is similar. Founders pitching me 2010-era org charts (VP Eng, VP Product, full exec team at Series A) are usually the ones building the wrong thing. The 2026 winners are running 8-person teams with agents doing what used to take 30.
@garrytan hot take: the most exciting 2026 businesses are re-building open source versions of products from the early 2000s eg
Firebase -> Supabase
Google Docs -> https://t.co/mjfnepl4Vx
Skype -> Jitsi
Calendly -> https://t.co/LHWJ6lVo0c
The obvious is so simple in hindsight yet obscured till you see it.
AO solves a lot of problems so elegantly, once you see it, it's an "Ah ha, of course, that is precisely how it should be"
Yet everyone else is still struggling with moniliths
AO devices are like *nix tools: Simple, infinitely remixable, and open for all to build.
This is why you can build things like trust-minimized oracles in a single HTTP URL (yes, really [1]).
But building devices used to suck.
Yesterday, the device forge flipped that. Details👇
Your AI agent can read your email, deploy your code, and charge your credit card. What happens when it decides to do something you didn't intend? That's the gatekeeper problem.
The same access that makes an agent useful also makes it dangerous. Most teams solve this by dumping an API key in a .env file and hoping for the best. Clawvisor takes a different approach. Every agent request needs a purpose.
Every purpose gets verified. Every execution gets audited. It's the permission architecture between your agent and your services; credential vaulting, scope verification, and a full audit trail, all in one layer. In this video, we break down why trust isn't architecture, how purpose verification catches scope mismatches, and why the companies running agents safely aren't the ones with the smartest models; they're the ones with the best permission infrastructure.
-> https://t.co/vFW5Oeo615
-> https://t.co/iFvCJFHrVJ