The interesting number in PrismML’s Bonsai 27B announcement isn’t the parameter count — it’s 3.9 GB. That’s the 1-bit build of a 27B-class model, small enough to sit in an iPhone 17 Pro’s memory budget. The ternary variant lands at 5.9 GB and runs on an everyday laptop. Both keep the full stack — reasoning, structured tool calls, vision, and multi-step agentic loops — instead of degrading to a toy once you quantize hard.

What matters operationally is where the model can live. A 27B model at 4-bit is ~18 GB, which means a server round-trip for every tool call in an agentic loop. Push effective precision down to 1.125 bits across the whole network — embeddings, attention, MLPs, and the LM head, not just the transformer blocks — and the model moves to the edge. The latency budget for a multi-step agent changes shape entirely when the loop isn’t crossing a network boundary between each step.

The claim I’d want to pressure-test is coherence across steps. Aggressive quantization tends to show up not as a headline benchmark drop but as drift — the agent loses the thread by step eight, tool arguments get subtly wrong, the vision grounding wobbles. A 27B model that stays sharp for a single-turn eval and falls apart over a long computer-use trajectory is a different product than one that holds. That’s the eval that decides whether “runs on a phone” is a demo or a deployment target.

The HN thread is already arguing about whether 1-bit weights can carry real reasoning or just pattern-match well enough to look like it. If the multi-step numbers hold up, the more disruptive question is for the inference-vendor economics: what’s your per-token pricing story when a capable agent runs locally for free?