DeepSeek's V4.1F has fewer active params than any prior flagship — compute scarcity biting
teortaxesTex · x · 2026-10-19
teortaxesTex notes that DeepSeek's V4.1F has fewer active parameters than any major DeepSeek model ever—less than V1-Coder (Nov 2023), V1, V2 (May 2024), V3 (Dec 2024), and V4 Pro (Apr 2026).
In the quoted tweet he argues that DeepSeek's strongest model today has fewer total and far fewer active params than V3 from Dec 2024; MiniMax shows the same pattern (M1, Jun 2025, was 456B total / 46B active). He reads this as compute scarcity biting: resources are being reallocated to RL rather than larger dense models.
More from AGI Musings
- Vicki Boykis on How to Learn ML in 2026: a 1-hour pragmatist's guide — yacinelearning · 2026-10-19
- Delta, JPMorgan, Amazon move to block AI agents as the web pushes back — AccBalanced · 2026-10-19
- Paras Chopra: startups must find what compounds against huge token budgets — paraschopra · 2026-10-19
- Anthropic: Engineers' New Role Is Saying "No" to Claude — But Who Trains the Next Generation? — sqcai · 2026-10-19
- Boaz Barak at Princeton: "Alignment Is Not a Solved Problem" — KatieMc___ · 2026-10-19
- Economist John Horton Calls 'Stochastic Parrots' a Midwit Concept; Mollick Slams AI-Can't-Manage Slide — soumitrashukla9 · 2026-10-19