Have you ever stared at a phone in your pocket and thought, this thing is mostly idle? I have. More than once. That quiet little computer sits there between messages, maps, and the odd photo, while the heavy lifting of artificial intelligence still happens in warehouses packed with servers. That split always felt a bit strange. The device in your hand is powerful enough for plenty of quick choices. It just rarely gets asked to make them for anyone else.
What Changed When Laya Landed On Acurast
Acurast has now put an open source decision model called Laya onto its decentralized smartphone compute network. The pitch is simple, and I find it more interesting than another chatbot launch. Instead of generating long replies, Laya looks at a defined state, weighs a set of possible actions, and picks one. Classification. Routing. A move in a game. A yes or a no. That is the job.
The work runs on Android phones that have joined the network. Not on a rented rack in a distant region. Decisions are described as landing in roughly 0.2 to 1 second, using phone CPUs rather than dedicated graphics cards. If that number holds up under messy real traffic, it matters. Plenty of products do not need a novel. They need a fast structured answer.
By running Laya on Acurast, we’ve changed that. We are proving that System 1 AI can run securely and at scale on hardware everyone already owns, providing a truly open, decentralized alternative to centralized cloud lock-in.
– Acurast founder Alessandro De Carli
I like that phrase, System 1. It is borrowed from the old idea that some thinking is fast and patterned, while other thinking is slow and verbal. Most public AI talk still orbits the slow, wordy kind. Laya is built for the other pile. Given options, score them. Given a range, pick a number. Given a claim, say true or false. The public repository frames those three formats pretty clearly.
Why A Decision Model Is Not Just A Smaller Chatbot
People lump every model into the same bucket. That is lazy, and it hides the product design. A conversational system keeps producing tokens until it decides to stop. A non autoregressive decision model does not stroll through a sentence. It evaluates and exits. That architecture is the point.
Laya comes from Convai Innovations under an Apache 2.0 license. Weights are public. Code is public. There is a 421 million parameter English checkpoint based on ModernBERT large, a multilingual version around 322 million parameters covering more than 100 languages, and a separate checkpoint aimed at typed workflows such as customer service, invoices, and security incidents. That mix is more useful than a single giant file that tries to be everything.
In my experience, teams waste money sending every tiny judgment to a frontier model. Spam or not. Headline category. Is this prompt trying to hijack an assistant? Those questions do not need poetry. They need a cheap, repeatable choice with a trail you can check later.
- Choice tasks pick among a defined set of options
- Scoring tasks assign a value inside a stated range
- Boolean tasks return a clear true or false
Keep the option list short. The project itself warns that general checkpoints wobble on some zero shot typed tests and that choice questions should stay under roughly twenty options. Recalibrate with your own data. That is not a footnote. That is the difference between a demo and a product.
The Network Behind The Demo
Acurast is not new to the idea of squeezing work out of phones people already own. Earlier fundraising supported a decentralized cloud built on consumer hardware. The company now says more than 280,000 smartphones across over 175 countries have joined, with on chain transactions past 918 million. Those figures come from the project. Treat them as company reported, not as an audit you can wave in a board meeting.
Still, the direction is hard to ignore. A year earlier the same network was described with tens of thousands of devices and a few hundred million transactions. Growth like that, if real, changes the conversation from novelty to capacity planning.
Work is handed out dynamically. Results arrive with cryptographic proof. Operators earn ACU, the native token, for offering processing time. Modern phones also ship with Trusted Execution Environments. Acurast says those enclaves isolate jobs and protect data while a decision runs. That claim is central. Without isolation, handing strangers a classification job on a personal device is a non starter.
What The Live Tests Actually Look Like
Demos can be silly and still teach you something. Laya has been pointed at Snake and Tetris. Move left. Rotate. Drop the piece. A Doom style test asks for movement, aiming, shooting, item pickup, and exploration. None of that will replace a studio AI team tomorrow morning. It does show a model making repeated micro choices under time pressure on ordinary silicon.
The more practical suite sits closer to software people already ship.
- Sort mail into inbox, spam, or phishing
- Label headlines as news, satire, clickbait, or manipulation
- Flag prompt injection attempts aimed at assistants
- Spot toxic lines in a live chat stream
Users can paste their own letters, headlines, and prompts and watch a decision come back from the mobile edge. That public loop is smart marketing. It is also a quiet stress test. If the network stalls when strangers hammer it with junk text, you will hear about it.
Perhaps the most interesting aspect is not the games. It is moderation and security sitting next to entertainment on the same compute fabric. Same phones. Same proofs. Different stakes.
Phones Without GPUs Can Still Be Fast Enough
We have been trained to believe serious inference needs a graphics card and a cooling budget. For giant generative models, that is still true. For a few hundred million parameters doing structured decisions, the story bends.
Acurast says these jobs run on smartphone CPUs. Latency in a fraction of a second is not magic if the model is small and the output is a label rather than a paragraph. I have found that people underestimate how much product work lives in that band. Routing a ticket. Scoring a risk flag. Choosing the next hint in a learning app. Those loops happen all day.
| Workload | Typical output | Why phones can fit |
| Classification | One label | Short context, fixed classes |
| Moderation | Allow, review, block | Repeated tiny judgments |
| Game policy | Next action | Bounded action space |
| Security triage | Safe or suspicious | Pattern match, not essay |
Does this replace a data center for training? Of course not. Training still likes dense clusters. Inference for narrow tasks is the opening. That opening is wide if you stop treating every request like a conversation.
Open Weights Change Who Gets To Ship
When weights sit behind a private API, you rent access and you inherit someone else’s rate limits. When weights are public under Apache 2.0, a developer can specialize the model, pin a version, and run it where the network allows. That is a different power map.
It also puts responsibility back on the deployer. Open does not mean finished. The maintainers already say the general model struggles on some typed zero shot tests. Specialize it. Keep choices tight. Calibrate probabilities with data from your actual users, not from a pretty README.
I would rather see that honesty than another launch post that pretends a mid size encoder is a universal oracle. Tools that admit their edges age better.
Decentralized Compute Is Not A Slogan If Proofs Matter
Plenty of projects talk about unused devices. Fewer talk about verification. Acurast’s story leans on cryptographic proof of the result and on hardware isolation. Those two pieces decide whether this is a science fair or infrastructure.
Think about a moderation decision that later gets challenged. You want more than a screenshot. You want a record that a given job ran inside an expected environment and returned a given output. Blockchain adjacent networks like to promise that. Delivery is the hard part, especially when devices drop off Wi Fi, batteries die, and users close an app.
Dynamic scheduling has to absorb that mess. Phones leave. Phones return. A job that needed 400 milliseconds cannot wait for a device that went into a subway tunnel. The network either retries fast or the product feels broken. That operational grit will matter more than any single demo clip.
Where This Sits Against Other Mobile AI Bets
Acurast is not the only group trying to turn consumer devices into a compute layer. Other efforts have pushed phones that run models locally and still let owners contribute spare cycles. The pattern is the same even when the brands differ. Hardware is already in pockets. Cloud invoices keep rising. Privacy rules keep tightening. Something has to give.
The difference here is the workload choice. Long form generation on a midrange handset still hurts. A compact decision model is a better first tenant. Start with jobs that finish quickly and produce structured output. Then argue about poetry later.
This deployment proves that decision-oriented AI can run cheaply, verifiably, and without a data center in sight on a decentralized smartphone network today.
– Alessandro De Carli
That sentence is a thesis. Cheap. Verifiable. No warehouse in the loop. If even two of those three hold under load, developers will listen. If only the demo holds, it stays a press cycle.
Money, Tokens, And The Awkward Incentive Question
Operators get paid in ACU for capacity. That is the obvious loop. It is also the fragile one. Token rewards can attract devices that vanish when prices sag. They can also attract junk hardware that passes a check and then underperforms.
I have watched enough networks to be wary of headline device counts. What you want is reliable capacity at peak hours in the regions your users actually live in. A phone in a drawer at 3 percent battery is not a node. A phone on a desk with stable power and a TEE that works is a node.
Earlier, the project raised several million dollars to grow the smartphone cloud idea. Funding buys time. It does not buy trust from a security team that has to sign off on sending customer text to unknown handsets. Proofs and enclaves have to survive that review, not just a keynote.
Limits You Should Read Before You Get Excited
Let me be blunt. A 421 million parameter English model will not reason through a messy legal dispute. It will not write a launch plan. It will not replace a large generative stack for open ended work. If you ask it to do those things, you will be disappointed and you will blame the network for a model mismatch.
Zero shot typed decisions can be weak. Too many choices dilute quality. Probabilities need calibration. Multilingual coverage is broad on paper and uneven in the wild, as it always is. Android only, in this telling, leaves a large slice of devices out of the current story.
Battery life is the quiet villain. Users will unenroll if their phone gets warm in a meeting. Any serious deployment has to throttle, schedule around charging, and stay invisible. Invisible is harder than impressive.
Practical rule of thumb I keep coming back to: Small model Tight action space Fresh calibration data Proof you can show a skeptic
What Developers Can Actually Try
Because the software and weights are public, a builder can start with a narrow slice. Email triage for a community inbox. Headline sorting for a research desk. A first pass on jailbreak-like prompts before a heavier model ever sees the text. Those are boring jobs. Boring jobs pay the bills.
The company has opened demonstrations so people can submit their own inputs. That is a low friction way to feel the latency with your thumbs instead of a slide deck. Feel matters. A 700 millisecond label that arrives with a proof feels different from a 4 second paragraph that arrives from a vendor you cannot inspect.
If I were advising a small team, I would not rip out an existing cloud classifier on day one. I would shadow it. Send the same payload both ways. Compare labels. Compare cost. Compare failure modes when a phone drops. Keep the cloud as a fallback until the edge path is dull and reliable.
Privacy Is The Real Product Feature
Central clouds are convenient and they concentrate risk. A breach or a subpoena hits a warehouse of other people’s text. Edge processing on isolated hardware is not automatically private, but it changes the shape of the risk. Data can be processed closer to a user and discarded faster if the application is designed that way.
That design is not guaranteed by a press line. Someone still has to decide what leaves the device, what sits in an enclave, and what gets written to a chain. Those choices are product work. Skip them and you have a distributed leak instead of a distributed cloud.
Still, the direction is healthier than pretending every classification needs to live next to a foundation model’s logs. Some text should never take that trip.
The Broader Shift Toward Spare Consumer Hardware
Look around a city. Laptops asleep. Consoles idle. Phones charging on kitchen counters. For years the industry answered demand by building more specialized buildings. That worked. It also created a bottleneck around chips, power, and a handful of providers.
Using hardware that already exists will not erase that bottleneck. It can nibble at the jobs that never needed the fancy building. Real time decisions are a clean nibble. They are frequent. They are small. They hate waiting on a round trip if a local or near local node can answer.
Is every phone a perfect worker? No. Heterogeneous fleets are ugly. Some chips are old. Some OS versions are awkward. Scheduling across that zoo is an engineering slog. The slog is the moat, if anyone finishes it.
A Sober Read On Scale Claims
Two hundred and eighty thousand devices sounds large until you ask how many are online at once, how many pass integrity checks, and how many sit in the latency budget you actually need. Transaction counts can climb for reasons that have little to do with AI inference. Do not confuse ledger activity with model throughput.
Ask for occupancy. Ask for p95 latency by region. Ask what happens when a popular game demo spikes. Those questions separate a narrative from a platform. I would rather be the person asking them than the person quoting only the biggest number on the page.
Why This Story Stuck With Me
I am tired of launches that only add more words. More tokens. Longer answers. Softer tone. The unglamorous work of choosing among options is where software actually runs. If a compact open model can do that on phones people already paid for, the economics of a lot of quiet features change.
Will Laya become a household name? Probably not. Decision models rarely do. The winners in this category tend to disappear into products. That would be a compliment.
The test over the next stretch is ordinary. Does a developer keep using it after the novelty fades? Does a phone owner keep the app installed after the first warm battery? Does a security reviewer sign the memo? Those are the three doors. All three have to open.
What To Watch Next
Watch for specialized checkpoints that beat the general English model on one boring task. Watch for published latency distributions, not just a range that sounds friendly. Watch whether TEE backed proofs become something outside teams can independently check. Watch device churn when token prices move.
Also watch copycats. Once you show Snake running on a stranger’s phone with a decision model, someone else will try a similar stunt with a different checkpoint. That is healthy. It means the idea is bigger than one brand.
If the category works, the long term picture is a split stack. Heavy generation stays in clusters. Fast structured judgment fans out across edges. Users feel less lock-in. Builders feel more choice. Operators of idle phones pick up a little yield without buying a server they do not want.
That split is not guaranteed. It is plausible. Plausible is enough to keep paying attention, especially when the hardware bill for intelligence keeps climbing and the unused computer in your pocket keeps sitting there, waiting for a job that actually fits.
So here is where I land. Laya on Acurast is not the end of centralized AI. It is a reminder that not every decision needed the cathedral. Some needed a phone, a tight question, and a proof you can hold up when someone asks how the answer arrived. That is a smaller story than a talking machine. It might be the more durable one.
