The sub-bit threshold: when 27B stops needing a server

The threshold just moved. The size class of model that ran cloud chatbots for the last two years now fits inside a personal device's working memory. That isn't a size-class record. It's the moment the cost curve of "intelligence" stopped being about data centers and started being about the phone in a pocket.

Call it the sub-bit threshold: when effective bits per weight fall low enough — 1.125, in PrismML's Bonsai 27B — that the size class the cloud depends on becomes an app. The mechanism is mechanical, not magical. Compress every weight to one of two values, keep a small FP16 scaling factor per group, and 27 billion parameters that needed a GPU rack become 3.9 GB of memory. Add speculative decoding and a 262K-token context, and the laptop-class ceiling quietly relocates to the device.

The reading people will reach for is "AI is getting cheaper." The stronger one — a reading the announcement implies but does not state — is that control is getting cheaper. Every parameter that no longer needs a server is a parameter that no longer needs anyone's permission, throttling, or data policy. The on-device price collapse doesn't shrink the model — it shrinks the gatekeeper. PrismML's Bonsai release didn't just ship a small model. It shipped the gate at a size where personal hardware can clear it.

The tradeoffs haven't changed: battery, thermals, latency, and a 90% self-reported quality figure from the vendor's own 15-benchmark suite. The floor under those tradeoffs just moved from a server to a pocket.