0.6B reads 27B

A 0.6 billion parameter model runs in this tab, on your hardware. When you send it a message it writes a reply, and the same model also searches a gallery of photographs. Not two models and not a call to an embedding service: your message is read once more by the same frozen weights, and a 2.1 MB linear head turns that hidden state into the vector that searches.

Those photographs were encoded once, months ago, by a 27 billion parameter model on a datacenter GPU. That computation is finished and it is in the file: what you download is its output, about a kilobyte per photograph, not the weights that produced it. Nothing of the large model runs while you search, here or anywhere else.

Language models running in a browser are not new, and cached image embeddings are ordinary. The read-out is neither. Text encoded by the small model in this tab lands in the same space as images encoded by a model forty-five times its size, because both sides are linear read-outs fitted on frozen backbones. Cross-scale retrieval where the large half is a finished file.

Nothing you type leaves the tab. No inference server, no API key, no telemetry: eight megabytes of static files, computed entirely on your device. It runs on the CPU through WebAssembly, so there is no WebGPU requirement and nothing to install, which is why a phone can run it from a link and keep working offline after the first visit.

not loaded

Try one, and watch the scores as much as the pictures:

Each turn is answered with the conversation so far in front of it, and searched with only the message you just sent. The pictures come from reading that message through the same weights that write the reply, and that read finishes long before the sentence does.

Ask something with a scene in it and the scores spread out. Ask the model about itself and you will keep seeing the same photographs: a question with no visual content projects to near the middle of the gallery, where the closest images are simply the ones nearest the average of all of them.

The line between those two is not abstract against concrete, it is whether the words name anything physical. Semiosis inside your frozen state is as abstract as a sentence gets, and it retrieves bodies curled inside suitcases and drawers, because frozen and inside are things a photograph can be.

Measured against these 123,287 images, a scene runs 0.47 to 0.54 with a spread of about 0.06 across the twelve results. The frozen-state question and a question about the model itself both sit near 0.41 across roughly 0.05, and the second of them actually scores higher. So the top number alone will not tell you which is which. What separates them is that one of them keeps returning the same photographs whatever you ask.

Why two reads rather than one? Because the head is fitted on plain text, and the chat format the reply needs is a different distribution. Feeding the templated prompt to the head instead of your bare message moves the median rank of the right photograph from 187 to 41,014 out of 123,287, and the two vectors have a cosine of 0.10. Sharing a single pass would be tidier and it does not work. Same lesson as the four kilobytes below: a linear read-out is fitted to one distribution of states and does not survive being handed another.

There is a coincidence worth noticing. Ask it what it is thinking about and it will say something like the world around me, and the people I've met. Meanwhile the search, having nothing visual to work with, falls to the centre of the gallery. We went and looked at what is there: of the eight images nearest the centre, seven are people and five of those are single portraits facing the camera. The sentence and the pictures are produced by completely different machinery, and with nothing to go on both fall back to the same thing.

Steering

An axis is the difference between two sets of sentences, measured in the space the search happens in. Adding it to your query reorders the gallery. It costs one vector addition and about two kilobytes; the model is not touched and nothing is re-encoded.

minus

strength 0.50

The table reports retention: how much of your query's own result survives. Turning the strength up looks more dramatic and is worse. Past about 1.0 the axis has replaced your query and every search returns the same pictures, which is why the calibrated default is 0.5. The random control is here so you can check that a real axis does something a matched piece of noise does not.

What this does not do

Steering here acts on the search query, not on the model's generation. The research programme also measured a direction in the residual stream that moves a model's refusal disposition, in both directions, and the finding worth leading with is the restoring one: a calibrated two-kilobyte vector puts refusal back into a model that had it removed. We also measured that the removal edit perturbs the model about eighteen times less than switching quantizers does. Those belong in the papers. They are a different artifact with a different risk profile, so this page exposes no hook into generation at all, and the reply above is produced by the model unmodified.