New MoE base claims win over Qwen3.8-Max with 50% fewer tokens, met with skepticism
teortaxesTex · x · 2026-11-16
A leak claims a new MoE-architecture base model outperforms Qwen3.8-Max with up to 50% fewer tokens and over 60% lower cost per task. Commentator teortaxesTex is skeptical: those numbers are what you'd ship in a next-gen FLASH-tier model, not a flagship, signaling capability gains have hit a wall. He also ranks the field: Alibaba is "out for the count," DeepSeek is the top small lab, Tencent the top big lab.
More from Companies & People
- Anthropic alumnus Jacob Coxon admits keeping OpenAI equity, calls it "blood money" — ns123abc · 2026-11-16
- Free online textbook 'Bayes, AI and Deep Learning' spans probability to agents — burkov · 2026-11-16
- Why Firms Don't Get Value From AI: Use It as a Proactive User of Firm Context — Afinetheorem · 2026-11-16
- Ex-OpenAI researcher: America can't produce a pragmatic open-source AI lab, nonprofits won't cut it — suchenzang · 2026-11-16
- Matsuzaki backs Socher in spat over paper-replication work vs. real AI invention — arankomatsuzaki · 2026-11-16
- WSJ alleges Huang built Southeast Asia data center route giving China backdoor to Nvidia chips — acmoytoy · 2026-11-16