They Squeezed a Model That Can See Onto a Phone. The Box Watching Your Pump Can See Now Too.

Todd Deshane · July 2026 · 6 min read

Every few weeks a number comes out that quietly moves the floor under my business. This week's number is 3.9 gigabytes.

A Caltech spinout called PrismML released a model named Bonsai 27B. The "27B" means 27 billion parameters, which is a real, capable model, not a toy. The interesting part is that they compressed it down to a 3.9GB file and ran the whole thing on an iPhone. Not a phone talking to a server. The model itself, living on the device, answering at about 11 words a second. They did it with an aggressive trick called 1-bit quantization, and the headline claim is that it kept about 90 percent of the full model's capability across 15 different tests. An eightfold size cut, and it barely got dumber.

That alone is a nice story about phones. But it's not why I stopped what I was doing. I stopped because of one word in the spec sheet that most of the coverage skipped: multimodal.

The word that matters is "multimodal"

Multimodal means the model can look at a picture, not just read text. You can hand Bonsai an image and it will reason about what's in it. And now that model runs on a cheap local device with no internet connection.

For most people that's a convenience. For someone who puts monitors on equipment in basements and boiler rooms, it's a door opening. Because reading the physical world has always split into two jobs, and until now only one of them could happen locally on cheap hardware.

The first job is watching a signal, vibration, current draw, temperature, and noticing when it drifts. Small models have done that on tiny edge boards for years. That's the core of every monitor I install. The second job is looking, reading a gauge face, a nameplate, a pilot light, a puddle on the floor, and understanding what you're seeing. That job needed a real vision model, and a real vision model meant sending the picture up to a cloud API, paying per call, and hoping the connection held.

A model that can look at a gauge and reason about it now fits in under four gigabytes on a local box. The "looking" half of monitoring just stopped requiring the cloud. The gateway on the wall can read the dial, read the nameplate, and correlate what it sees with the vibration trend it's already tracking, without a single API call leaving the building.

Two years ago, if I wanted a monitor to read as well as watch, I had two bad choices. Bolt on more single-purpose sensors, or pipe camera frames to a cloud vision service and eat the latency, the monthly bill, and the privacy problem of a customer's boiler room streaming to someone else's servers. Bonsai 27B is the week the third choice became real: the looking happens on the same $50-to-few-hundred-dollar box that's already doing the watching.

What this actually looks like on a machine

I've written before that every gauge is a sensor. An analog dial, a temperature readout, a little sight glass showing oil level, those are all data, they just happen to be printed instead of wired. For years the way to get that data was a person walking by with a clipboard, or a camera shipping the image somewhere smart enough to read it.

Put an on-device multimodal model behind a five-dollar camera and the picture changes. The gauge gets read on the wall, every few minutes, forever. Better than that, the model can hold two things at once. It can see that the pressure dial reads high and know that the vibration signature started climbing this morning, and treat those as one story instead of two disconnected alerts. The nameplate photo you took during install stops being a filing-cabinet JPEG and becomes something the box can actually reference: model number, service ratings, the specific pump this is.

None of that needs a data center. That's the whole point. The intelligence that used to justify a cloud subscription now runs next to the machine, on hardware that costs less than the service call it prevents.

You were probably told that "smart" monitoring, the kind that can read a display and understand what it's looking at, requires the cloud. That was true last year. It stopped being true this week. A capable model that sees and reasons now runs on a local board with no connection, no per-image fee, and nothing leaving your property. The premium half of monitoring just became the boring, on-premise, flat-rate half.

This is the curve, and it only goes one way

Bonsai 27B isn't a one-off. It's the same curve I've been riding since I put the first monitor on a sump pump. The capable models keep getting smaller and cheaper to run locally, and the cheap local hardware keeps getting good enough to run them. A 27-billion-parameter model that can see, running on a phone, is this week's mile marker. Next quarter it's a $60 edge board. The direction never reverses.

What that means for a building owner is simple and a little unglamorous. The gap between "the fancy monitoring the big facilities get" and "what fits on your pump for a flat monthly fee" keeps closing, and it closes on your side. You are not waiting for this technology. The expensive, cloud-dependent version of it is what a data center bought last year. The cheap, local, already-shipping version is what I clamp onto a machine on a Tuesday.

What to take from this if you own a building, not a phone

Last week I said a single cheap camera and a small model beat the expensive sensor array. This week sharpens the other half: that small model can now see, and it does it on the wall in your basement instead of on a server you rent. I sell the finished, boring version of that, one cheap sensor, a local box that both watches and looks, on your machine, today.

The box on the wall can watch your machine and read its gauges. No cloud, no per-image bill.

Each critical asset, your compressor, your boiler, your pump, gets one off-the-shelf sensor and a small model on a local device. It learns that machine's normal, reads the gauges and nameplate it can see, catches drift three to six weeks before a failure, and hands it to a person who knows the equipment. Nothing leaves the building. $99 to $199 per month, hardware under $3,000.

See how it works

Sources: PrismML, Bonsai 27B (Qwen3.6 27B base compressed via 1-bit quantization to a 3.9GB binary variant and a 5.9GB ternary variant; multimodal, accepts images and text; ~11 tok/s on iPhone 17 Pro, up to 87 tok/s on M5 Max; retains ~90% of full-precision performance across 15 benchmarks; Apple reportedly evaluating the compression tech), released July 14 2026, via prismml.com/news, 9to5Mac, and DeepLearning.AI Data Points (July 15 2026). Edge predictive-maintenance economics (local inference <200ms; 50–70% reduction in unplanned outages within the first quarter; 12–18% lower overall maintenance cost) via IndustryWeek and EE Times, 2026. Field deployments at The Intersecto Watertown sump-pump site and Northampton 40-device building. Companion brief: /Users/tdeshane/lobster/research/physical-ai-brief-2026-07-18.md.