Ant Group Zhou Jun AICon Speech Shifts Focus From Token Quantity to Token Density for Trillion-Parameter Models

Ant Group vice president Zhou Jun addressed efficiency limits at the AICon conference. He noted that a trillion-parameter model running for 15 minutes incurs compute costs equal to one Tesla vehicle.

This comparison set the core problem for agent-era systems. Efficiency must improve or scaling will stall.

The speech proposed moving from raw token counts to higher token density. The goal is to deliver more useful computation per token generated.

The company described a hybrid linear attention design. It combines seven parts Lightning Attention with one part MLA. The mix reduces 256K context costs from exponential growth to linear scaling.

Computational resources now go toward reasoning instead of basic context handling. The architecture keeps long-context windows manageable at trillion-parameter scale.

Ant Group also introduced the Kpop algorithm. It separates tool-calling tokens from natural language tokens. This separation allows pruning of low-value reasoning chains and self-distillation steps.

Output token volume dropped roughly four times on tested tasks while benchmark scores held steady. LongBench and BFCL results showed gains for hundred-billion-parameter agents against some larger baselines.

Smaller flash models reached 2.4 times higher throughput. Five-turn dialogue costs fell more than ten times in internal tests.

The approach pressures pure scaling strategies that rely on ever-larger context windows. Competitors continue expanding raw context length without similar density gains.

Critics note that the reported benchmarks come from Ant Group evaluations. Independent large-scale verification remains limited.

The hybrid attention method still requires careful tuning of the seven-to-one ratio. Any deviation can reintroduce non-linear cost spikes.

Future signals to monitor include open releases of the Kpop algorithm, third-party reproduction of the linear-cost claims on LongBench, and adoption rates inside production agent platforms over the next quarter.

Developers building long-context agents should examine density metrics alongside raw context length when choosing base models.