8 atoms · 跨 5 天 · 首见 2026-04-16 · 最近 2026-06-11
三色: 🟦 fact 5 · 🟥 take 3 · stance ▲1/▼1/◆2
来源: X 8
时态: fresh:6 · aging:2
标签: 好数字:5 · 好观点:2 · 好思考:1 · 好问题:1
[老化中] · MTP modification architecture1. Uses 4 decoder layers with 3 SWA + 1 GA
2. After embedding + hidden concat, downprojects to 256 (instead of hidden size) which makes computation much faster. Before LM head, uprojects back to hidden size.
3. Much more efficient LM head in the MTP by selecting clusters of logits!_
tags: 好观点·好思考 · → daily
[老化中] · 面临Qwen3.6-35B-A3B和Qwen3.5-27B两面夹击qty=45 rightqty=5 wrongqty=218 tok/s · date=benchmarkqty=15.1s · date=benchmarkqty=1.5x fasterfact · 正确事实数 · → dailyfact · 错误事实数 · → dailyfact · 输出速度 · → dailyfact · 总推理时间 · → dailynarrative · 与V4-Pro的性能差距 · → dailyfact · 本地模型推理速度提升 · → dailynarrative · MTP modification architecture · → dailynarrative · 面临Qwen3.6-35B-A3B和Qwen3.5-27B两面夹击 · → daily