Helix 2.5 skips home-training and goes straight to 30 real apartments with zero in-domain data

Corey Lynch unveiled Helix 2.5 over the weekend — a full-body model that executes three distinct behaviors across 30 real apartments without a single training frame collected inside any of them. The model simply "got it" zero-shot, meaning no environment-specific examples whatsoever.

The engine behind the leap is Index, a data platform that as of last Saturday was generating 35 minutes of training per second from 90 thousand weekly active users. Those numbers have since climbed past 100 thousand weekly users and 50 minutes per second. Index doesn't just deliver volume; it supplies a continuously updated stream of full-body demonstrations that feeds the model continuously.

The most telling figure comes from a straight ablation: same model, same task data, same evaluation. Without Index pre-training, zero-shot success sits at 8 percent. With it, 56 percent. That 48-percentage-point gap shows Index transfers entire whole-body behaviors zero-shot, not merely better generic representations.

The team also reports clear scaling signals. Doubling Index data improved next-action prediction consistently enough to forecast the loss of the largest run to four decimal places before the run even started. Lynch says massive pre-training is starting to feel fundamentally different — not just "more data," but predictable dynamics.

The company has committed $3.5 billion in compute for the next Helix version, and Index keeps growing daily. The full Helix 2.5 report is available now. The roadmap is straightforward: increase capacity, extend the training window, and see whether the jump from 8 percent to 56 percent is only the beginning of the curve.