以下部分引用 AI 总结,现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
JetBrains Mellum 2
13 atoms · 跨 1 天 · 首见 2026-06-01 · 最近 2026-06-01
三色: 🟦 fact 13 · 🟥 take 0 · stance ▲0/▼0/◆0
来源: X 13
时态: fresh:13
标签: 好数字:13
事实 (fact) (13)
- 2026-06-01 · total parameters
12B total
tags: 好数字 · → daily
value: qty=12B
- 2026-06-01 · active parameters
2.5B active
tags: 好数字 · → daily
value: qty=2.5B
- 2026-06-01 · number of experts
64 experts
tags: 好数字 · → daily
value: qty=64
- 2026-06-01 · active experts per inference
8 active
tags: 好数字 · → daily
value: qty=8
- 2026-06-01 · pretraining tokens
~10.65T pretraining tokens
tags: 好数字 · → daily
value: qty=~10.65T
- 2026-06-01 · pretraining Phase 1 data composition
Phase 1: 70% web, 23% code, 6% math
tags: 好数字 · → daily
value: qty=70% web, 23% code, 6% math
- 2026-06-01 · pretraining Phase 2 data composition
Phase 2: 44% web, 42% code, 14% math
tags: 好数字 · → daily
value: qty=44% web, 42% code, 14% math
- 2026-06-01 · pretraining Phase 3 data composition
Phase 3: 23% web, 59% code, 18% math
tags: 好数字 · → daily
value: qty=23% web, 59% code, 18% math
- 2026-06-01 · KV heads
4 KV heads with GQA
tags: 好数字 · → daily
value: qty=4
- 2026-06-01 · FIM scheduled at 50% early
50% across all data early
tags: 好数字 · → daily
value: qty=50%
- 2026-06-01 · FIM scheduled at 10% during quality uplift
10% during quality uplift
tags: 好数字 · → daily
value: qty=10%
- 2026-06-01 · raw code seen for 3 epochs
raw code seen for 3 epochs, ~958B tokens
tags: 好数字 · → daily
value: qty=3 epochs, ~958B tokens
- 2026-06-01 · max dataset repetition
no dataset repeated more than 4×
tags: 好数字 · → daily
value: qty=4×
时间轴 (近 20)
- 2026-06-01 ·
fact · total parameters · → daily
- 2026-06-01 ·
fact · active parameters · → daily
- 2026-06-01 ·
fact · number of experts · → daily
- 2026-06-01 ·
fact · active experts per inference · → daily
- 2026-06-01 ·
fact · pretraining tokens · → daily
- 2026-06-01 ·
fact · pretraining Phase 1 data composition · → daily
- 2026-06-01 ·
fact · pretraining Phase 2 data composition · → daily
- 2026-06-01 ·
fact · pretraining Phase 3 data composition · → daily
- 2026-06-01 ·
fact · KV heads · → daily
- 2026-06-01 ·
fact · FIM scheduled at 50% early · → daily
- 2026-06-01 ·
fact · FIM scheduled at 10% during quality uplift · → daily
- 2026-06-01 ·
fact · raw code seen for 3 epochs · → daily
- 2026-06-01 ·
fact · max dataset repetition · → daily
← 实体目录 · 系统日志