Opinion

Ant Group's Zhou Jun at AICon: from token count to token density, efficiency first for trillion-parameter models

Original title:演讲实录 | 从 Token 数量到 Token 密度:万亿参数规模下的高效智能体智能

AI summary

Ant Group Vice President Zhou Jun said at AICon that a trillion-parameter model's compute cost for every 15 minutes of operation is roughly equivalent to a Tesla, and efficiency is the most urgent problem to solve in the agent era. The team proposed shifting from "more tokens" to "higher token density," adopting a hybrid linear attention architecture of 7 parts Lightning Attention to 1 part MLA, bringing the cost of 256K long context from exponential to linear and devoting more compute to thinking. Through the Kpop algorithm to distinguish tool-call tokens from natural-language tokens, combined with chain-of-thought pruning and self-distillation, token output is reduced by about 4x without degrading capability. It shows significant gains on benchmarks such as LongBench and BFCL, with a hundred-billion-parameter model surpassing some larger models on agent tasks; the small model flash throughput reaches 2.4x, and five-turn conversation costs drop by more than 10x.

Key Takeaways

  • Ant Group Vice President Zhou Jun said at AICon that a trillion-parameter model's compute cost for…
  • The team proposed shifting from "more tokens" to "higher token density," adopting a hybrid linear…
  • Through the Kpop algorithm to distinguish tool-call tokens from natural-language tokens,combined…
  • It shows significant gains on benchmarks such as LongBench and BFCL,with a…
  • the small model flash throughput reaches 2.

Quick Facts

Published
Updated
Source
公众号:蚂蚁百灵(Ling)
Category
Opinion
AI score
64
Read the original Read the full article at the original source
Why it matters

蚂蚁百灵副总裁周俊这次分享,把大模型效率问题从零散优化推到了架构、训练、智能体协同设计的范式层面,7+1 混合注意力方案和 Kpop 算法对做模型的人是实质参考。

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: Ant Group's Zhou Jun at AICon: from token count to token density, efficiency first for trillion-parameter models. July 8, 2026. https://new.nat6.net/en/news/cmrbdsefa02yjihl17gg7mhx3/

View all →