Ant Group's Ling team shows how to train a 290B model without relying on expensive GPUs
aftahi_ai · x · 2026-11-17
Ant Group's Ling team explains how they trained a 290B-parameter model without fully relying on expensive GPUs, via smarter architecture, hardware optimization, and better training strategies. The attached loss curves compare loss-fixing variants (micro batch-size, NormHead, NormHead grad, correct loss), showing the fixed loss trains markedly lower and more stable — making large-scale AI far cheaper to build.
More from Infra
- Agent Substrate: open-source zero-trust runtime suspends idle agents, fits 10x more per hardware — Saboo_Shubham_ · 2026-11-17
- Practical tips to cut agent memory costs on OpenAI: cache TTL, Flex, fewer calls — gabrielchua · 2026-11-17
- xAI affiliate buys 96 acres near Memphis Colossus 2 for $5.5M as compute footprint expands — XFreeze · 2026-11-17
- HSBC offloads $300M troubled AI data center loan at steep discount — MarcoFigueroa · 2026-11-17
- Last Year We Built Megakernels. This Year We Deleted Them — simran_s_arora · 2026-11-17
- SemiAnalysis shows Trainium3 profiling stack is precise enough to render pixel-art videos — joey00072fp8 · 2026-11-17