ZAYA1-8B

14 atoms · 跨 5 天 · 首见 2026-05-06 · 最近 2026-05-23

三色: 🟦 fact 10 · 🟥 take 4 · stance ▲6/▼1/◆3

来源: X 14

时态: fresh:11 · aging:2 · stale:1

标签: 好数字:10 · 好信源:5 · 好观点:2 · 好思考:2 · 好问题:1

叙事 (narrative) (2)

• 2026-05-08 [老化中] · active parameters count
  The "4B active parameters" framing makes it sound efficient, but they're achieving those results by burning extreme amounts of test-time compute that other models don't need.
  tags: 好思考·好问题 · → daily
  value: qty=4B · unit=parameters
• 2026-05-08 [老化中] · compute efficiency relative to DeepSeek-R1
  In reality it's at most 6,3x compute, and likely less, given that Zaya has 37x fewer active parameters.
  tags: 好数字·好思考 · → daily
  value: qty=6.3x · direction=less · condition=at most

立场 (position) (1)

事实 (fact) (11)

• 2026-05-23 · 模型训练
  A U.S. startup trained an open 8B reasoning model, ZAYA1-8B, entirely on AMD.
  tags: 好信源·好数字 · → daily
  value: qty=700M active params
• 2026-05-20 · HMMT 2025 数学基准表现
  On the HMMT 2025 mathematics benchmark, ZAYA1-8B beat Claude 4.5 Sonnet, GPT-5-High, and DeepSeek-R1-0528 with roughly 671 billion total parameters, two orders of magnitude more than ZAYA1-8B.
  tags: 好数字·好信源 · → daily
  value: direction=beat
• 2026-05-06 · active params
  With <1B active params, it outperforms open-weight models many times its size on math and re
  tags: 好数字·好信源 · → daily
  value: qty=<1B
• 2026-05-20 · 智能密度
  two times the intelligence per active parameter of the next-best open-weight reasoning model.
  tags: 好数字·好观点 · → daily
  value: qty=2x
• 2026-05-06 · 活跃参数数量
  With <1B active params, it outperforms open-weight models many times its size on math and re
  tags: 好数字 · → daily
  value: qty=<1B
• 2026-05-23 · 模型性能
  And it’s already punching at giant-model math.
  tags: 好信源 · → daily
• 2026-05-08 · inference token usage per problem
  ZAYA1-8B uses up to 5.5M tokens PER PROBLEM to hit those benchmark scores.
  tags: 好数字 · → daily
  value: qty=5.5M · unit=tokens
• 2026-05-08 · 发布
  ZAYA1-8B (760M Active)
  tags: 好数字 · → daily
• 2026-05-07 · 训练硬件平台
  a super efficient open reasoning model trained on AMD Instinct MI300 GPUs
  tags: 好数字 · → daily
  value: qty=AMD Instinct MI300 GPUs
• 2026-05-06 · 活跃参数量
  With <1B active params
  tags: 好数字 · → daily
  value: qty=<1B
• 2026-05-06 · benchmark scores are very competitive
  tags: 好信源 · → daily

时间轴 (近 20)

• 2026-05-23 · fact · 模型训练 · → daily
• 2026-05-23 · fact · 模型性能 · → daily
• 2026-05-20 · fact · HMMT 2025 数学基准表现 · → daily
• 2026-05-20 · fact · 智能密度 · → daily
• 2026-05-08 · fact · inference token usage per problem · → daily
• 2026-05-08 · narrative · active parameters count · → daily
• 2026-05-08 · narrative · compute efficiency relative to DeepSeek-R1 · → daily
• 2026-05-08 · fact · 发布 · → daily
• 2026-05-07 · position · active parameters story · → daily
• 2026-05-07 · fact · 训练硬件平台 · → daily
• 2026-05-06 · fact · active params · → daily
• 2026-05-06 · fact · 活跃参数数量 · → daily
• 2026-05-06 · fact · 活跃参数量 · → daily
• 2026-05-06 · fact · benchmark scores are very competitive · → daily

← 实体目录 · 系统日志