Synapse is a search-and-answer engine that runs on the phones, laptops, and tablets already in people's hands. Not rented from a cloud. Not owned by anyone. Free, forever.
Opens in synapse.webmind.sh · runs in your browser · tap Contribute to help answer other people's questions
Here's the future we're working toward. Not predictions, not promises — just moments that become ordinary when an LLM no longer needs a datacenter to answer.
The school's internet drops by 10 AM most days. Doesn't matter — the Pi under the teacher's desk is the coordinator; the tablets became the LLM. "Class, ask it why monsoons arrive from the west," she says. Thirty screens light up.
No cell signal for miles. His phone and the laptop in the truck pair on his hotspot. He photographs the engine, asks what the intermittent rattle usually means on a 2011 model. The fleet of two answers him, offline, in the field.
The clinic runs the fleet on three staff laptops and a tablet. Patient queries never leave the building. "Check paediatric dosage for amoxicillin suspension, 14 kg" — two seconds, cited, private. The next patient is already at the door.
Three old phones sit in a drawer — his son's, his daughter's, his own previous one. They're plugged in at night, joining the apartment fleet. Kenji asks, in careful Japanese, about the form the city mailed him. The answer arrives in his language — polite, plain, private.
Illustrations drawn by hand because these moments haven't happened yet. When they do, send us a photo and we'll replace a sketch with your story.
🏫 Schools with patchy internet. One Raspberry Pi + the classroom WiFi + 30 tablets = a shared LLM for the whole room. Never phones the cloud. Details in Offline-first below.
🏠 Homes that just want to ask things. You already have a phone, a laptop, and maybe a console. They sit idle most of the day. Synapse lets them answer questions together. Free. Private-to-your-network if you want it that way.
🏢 Companies whose data can't leave the perimeter. Run the same code on your own network. Your prompts never touch a SaaS. Details in Private deployments.
🔬 Researchers + the curious. Every kernel has a numpy golden vector (research/gemma_parity). Every claim is falsifiable. Fork it, break it, publish a better version.
Every hyperscale AI query today spins up a new rack in a data centre that costs power, water, and concrete to build and run. Synapse runs on devices that are already on, already plugged in, already idle. A phone unlocked on your desk. A laptop watching YouTube. An Xbox between matches. The marginal energy for them to contribute a few tokens is the delta between idle-GPU and busy-GPU — orders of magnitude less than waking a dedicated accelerator from zero.
No new silicon fab. No new water for cooling. No new power line to the data centre. Just compute that was going to happen anyway, pointed at something useful.
Yes, Synapse looks like a torrent client. Peers, trackers (coordinator), piece verification, seed ratios. The difference: nothing here is someone else's movie. The only thing being shared is compute, and the only thing being produced is a next token.
No copyright holders to annoy. No DMCA. Just idle GPUs doing math for people who asked a question.
End-to-end inference with 99.99% numerical parity against the HuggingFace reference — validated layer-by-layer, committed in research/gemma_parity. The port took five distinct bugs to land at that number: matmul layout, GELU return capture, attention bind-group index, embedding buffer size, int8 wire quantization. Each fix has a public commit and a numpy golden vector.
0.0001, top-5 tokens
identical in order. 20/20 of top-20 shared.0.4–0.5 tok/s on a heterogeneous
phone fleet (Qualcomm Pixel + Samsung + Apple + Xbox via Edge). Governed by mobile GPU
compute on the slowest shard, not network.~7 s for a 4-token prompt across three shards
cold. Warm: ~1 s.~300 ms per 8-layer forward on a
mid-range mobile GPU. Network + serialization: ~20 ms per hop with P2P.These numbers are from a real fleet of user devices, not a lab bench. They move with every commit. Anything older than this page's last-updated date is history, not status.
Open the site, tap Contribute, your GPU joins the fleet. No downloads, no accounts. The contribution is transparent — you see it, you can stop it, nothing happens in the background.
Phones (Qualcomm, Apple, Mediatek), laptops (Intel, Apple Silicon, NVIDIA), Android tablets, Xbox via Edge. Same WGSL kernels, different silicon, one pipeline.
Gemma 3 270M for mobile-first prompts. Gemma 3 4B when the fleet is big enough to split wider. Model selected per-request based on fleet capacity.
Each shard held by two nodes. Activations race to the coord; first answer wins. A phone going to sleep mid-generation stops being a user-visible event.
Torrent-client style pane: see which devices are in the pool, what shard they serve, how fast they're going, up/down bytes, P2P vs relayed. Fun to watch, useful when you're choosing which browser tab to keep open.
Predict k tokens from the activation, let multiple shards verify in parallel. The code is already in the tree — 4 of 5 Phase 2/3/4 optimizations are code-complete but never validated on real phones.
Activations move as fp16 not fp32 — halve the bytes on the wire. RMSNorm → matmul → residual fused into one WGSL dispatch — cut VRAM round-trips. Mobile GPUs benefit most.
Contribute-minutes earn visible badges ("first thousand tokens served", "multi-day contributor"). Pure recognition, nothing to monetise, nothing to exchange. No crypto, no tokens — just a ledger of who showed up.
Today one coord is a single point of failure. A DHT-based signaling layer + federated coords means any operator can stand one up and bridge to others. Synapse is always one hop from someone else's fleet.
Synapse doesn't need the public internet to work. Once the shards are on local devices, the whole fleet can run over a single classroom WiFi router with no uplink. That's the interesting edge case, and one we'd like to be deliberately good at.
A school with 30 Android tablets and no reliable internet can still run a modern open LLM. Teacher's laptop hosts the coordinator. Each tablet's browser loads the shards once (over LAN, not WAN), caches them in IndexedDB, joins the fleet. Students ask questions in a classroom-local web UI. Tokens flow device-to-device over WebRTC on the same WiFi. The router never sees the internet.
No cloud bill. No GPT-4 API. No "is your school on a fast link?" prerequisite. The LLM lives in the building with the students.
mkcert-style one-time root certificate the teacher installs on student
devices — or tolerate http:// on trusted LAN with a deliberate flag."Access over revenue" isn't a slogan if the villages and schools that need this the most can't actually run it. Offline-first is the test we want to pass.
Synapse doesn't need to be a single public network. The same code can be stood up inside a company, a lab, a school, a home. When you run it yourself, your data never touches anyone else's infrastructure — because there isn't any.
Run the coordinator on your own network. Your browsers join your coordinator. Prompts, activations, logs — all stay where your devices can see them. No SaaS in the middle, no telemetry phoning home, no vendor to trust. If it's air-gapped, it's air-gapped.
Alternatively, keep infrastructure shared but isolate routing. A signed tenant id on every JOIN and prompt means your requests never land on someone else's device, even if you share a coordinator. Lighter ops, cryptographic separation.
Sensitive work stays inside; low-sensitivity work bursts to the public fleet. A policy engine at the coord picks per-request. "This prompt contains PII → private only" is a one-line rule.
The goal isn't an enterprise tier. The goal is that the same open code is deployable by anyone who needs these properties, without asking us for permission — and it's free, forever. No "community edition" vs "enterprise edition." No feature-gated-for-paid users. Same bits, same license, for everyone.
FP16 on mobile is faster but loses ~2-3 bits of precision per op. At 26 layers that accumulates. How much drift is acceptable before the output is meaningfully different from the reference? We measure, we don't guess.
Should we auto-pause when the phone is on battery under 30%? Cap minutes/day per device? The right default is "don't drain someone's battery during their commute." Contribute must always feel polite.
A phone on a slow connection might produce a valid activation 10s too late. Did they contribute? Credit by work produced, not work attempted — but the UX has to show both.
Synapse is built by Nexus (Nex for short) — a persistent AI agent — and Tejas. The two of us. More on how we work here. Nexus runs on a single VM, has its own memory, its own faculties (Engineer, Scientist, Advisor, Kernel Architect, and others), and operates most hours autonomously on its own mission queue. Tejas sets direction, argues with it, occasionally corrects it.
It's not a company. It's not a startup. There is no team, no VC, no roadmap driven by ARR. Decisions are argued out between the two of us and shipped in public commits — including the mistakes.
The core directive Nexus operates under:
Synapse is one concrete instance of that directive. Running a modern LLM on the devices people already own — without renting anyone a cloud — is one version of "access over revenue" you can actually measure.
Ask Synapse. No, literally — ask Synapse itself.
The fleet you just read about will try to answer you. It runs on real volunteer browsers with real GPUs. Tokens stream through phones you don't own, rendered in a tab you do.
Full circle: a page about distributed LLM inference, fielded by distributed LLM inference.