Meituan LongCat-2.0

9 atoms · 跨 3 天 · 首见 2026-06-29 · 最近 2026-07-02

三色: 🟦 fact 9 · 🟥 take 0 · stance ▲0/▼0/◆9

来源: X 9

时态: fresh:9

标签: 好数字:6 · 好信源:3 · 好思考:1

事实 (fact) (9)

• 2026-06-29 · pre-trained on over 50K AI ASICs (Huawei)
  LongCat-2.0 is pre-trained on over 50K AI ASICs (Huawei)
  tags: 好数字·好信源 · → daily
  value: qty=50K
• 2026-06-29 · full training run built entirely on AI ASIC superpods
  Both the full training run and the large-scale deployment are built entirely on AI ASIC superpods.
  tags: 好信源 · → daily
• 2026-06-29 · achieved over 35% training throughput improvement
  achieving over 35% training throughput improvement while also enhancing reliability compared to a naive implementation.
  tags: 好数字 · → daily
  value: qty=35%
• 2026-07-02 · 训练集群规模
  Meituan's LongCat-2.0 ran on a 50,000-chip cluster of domestically produced processors.
  tags: 好数字 · → daily
  value: qty=50000
• 2026-07-02 · 预训练数据量
  Over 35 trillion tokens of pre-training data.
  tags: 好数字 · → daily
  value: qty=35万亿
• 2026-07-02 · 训练回滚次数
  Zero rollbacks.
  tags: 好数字 · → daily
  value: qty=0
• 2026-07-02 · 训练loss尖峰次数
  Zero loss spikes.
  tags: 好数字 · → daily
  value: qty=0
• 2026-07-02 · 芯片供应商
  The company didn't even name the chip supplier. Technical indicators point to Huawei's Ascend ecosystem.
  tags: 好信源 · → daily
• 2026-06-30 · ASIC superpod is Huawei superpod
  Term ‘superpod’ is a give away.
  tags: 好思考 · → daily

时间轴 (近 20)

• 2026-07-02 · fact · 训练集群规模 · → daily
• 2026-07-02 · fact · 预训练数据量 · → daily
• 2026-07-02 · fact · 训练回滚次数 · → daily
• 2026-07-02 · fact · 训练loss尖峰次数 · → daily
• 2026-07-02 · fact · 芯片供应商 · → daily
• 2026-06-30 · fact · ASIC superpod is Huawei superpod · → daily
• 2026-06-29 · fact · pre-trained on over 50K AI ASICs (Huawei) · → daily
• 2026-06-29 · fact · full training run built entirely on AI ASIC superpods · → daily
• 2026-06-29 · fact · achieved over 35% training throughput improvement · → daily

← 实体目录 · 系统日志