Liquid AI has put out two open-weight models that decline to do the thing everyone now expects models to do. They don’t write. d1-3B reads text and images. d1-omni-600M reads text with an image, or text with audio. Each returns a typed, calibrated answer in a single forward pass, with zero output tokens.
That reads like a feature list with something missing. Read it again and it’s closer to a bet: that a large share of what we currently point language models at isn’t language.
Most “generation” is classification in a costume
Think about what actually runs in production. Is this support ticket urgent. Does this frame contain a person. Is this audio clip a gunshot or a door slamming. Which of six routes does this request take. None of that needs prose. We generate prose anyway, because the interface the industry standardized on is a token stream, so we ask for JSON, parse it, retry when it comes back malformed, and bolt on a confidence field while pretending the number means something.
The decode loop is where the cost lives. Every output token is another forward pass. In a datacenter that’s a line item. On a camera, a sensor hub, a car, a headset, it’s the difference between a model that ships and one that doesn’t. Removing output entirely removes the variable-length tail from your latency budget. One pass in, answer out.
The claim worth poking at
“Calibrated” is carrying a lot of weight in this announcement, and it’s the first thing I’d try to break. A classifier that says 0.8 and is right 80 percent of the time is genuinely useful, because you can set a threshold and escalate everything below it. A classifier that says 0.8 and is right 55 percent of the time is worse than nothing, because it invites trust it hasn’t earned. Calibration on a benchmark suite and calibration on your factory floor, your doorbell, your particular mix of accents, are not the same measurement.
The design has a real cost, too. No output tokens means no reasoning trace. When one of these models gets a decision wrong, there’s nothing to read. You get a label and a number. Teams used to scrolling a chain of thought to find where a pipeline went sideways will miss it, and yes, that comfort was always somewhat fictional, but the debugging habit is gone either way.
Open weights is the aggressive part
The weights being open is what makes this more than a product note. A 600M model you can quantize, profile, and run on hardware you own, with no per-call bill, goes straight at the “just hit an API” reflex that has quietly become the default architecture for edge products with no business depending on a network round trip. Latency and privacy arguments have been made for years. Pricing is what usually settles them.
What I don’t know yet is how these hold up outside Liquid’s own evaluations, and that gap only closes when people start posting reliability curves from their own data. One detail to plan around in the meantime: d1-omni-600M takes an image or audio, not both in the same pass.