Every major model of 2026, and the topics that carry them, plotted as one rotating structure. Read from 3,906 field signals; then anticipated forward twelve months.
Time runs left to right (Feb–Jun 2026); each row is a topic; each strut’s height is how much the field was talking about it that month. The whole thing sits under a geodesic frame. Drag to turn it; hover a strut for the count.
The frontier shipped roughly three to four significant models a month. Open weights (red-marked at left of each) arrived at, and often above, closed-model class.
Fastest-accelerating topics, early-window vs late-window.
Open-weight agentic models crossed Opus-4.5 class; the question flipped from “is the gap closing?” to “how do you run it locally?” Almost every other story this month is downstream of it.
Fuller designed for where the structure is heading, not where it stands. Extrapolating the struts:
The open/China curve is the second-steepest and closed the gap in one quarter. Expect open models at or above closed frontier class to be normal, and access, not capability, to become the contested axis.
Hardware and kernels rose almost as fast as releases. DGX Spark, single-GPU, M3 Ultra, NVFP4: the winning story is “run the frontier yourself.” Sovereignty and cost, not benchmarks, drive the next year.
RL and post-training is the single fastest riser. The gains come from the loop, agents that work for hours, terminal RL recipes, not from bigger base models. The harness is the product.
Physical-AI signal doubled into mid-year. Once language agents commoditize, embodied and world-model work is the frontier that is still visibly open. Watch it climb the structure.