Notable change: K3 diverges from predecessor on vision encoders
stochasticchasm · x · 2026-09-11
A developer notes that K3 shows a big difference from its predecessor on vision encoders, and also highlights the Adam/RMS-matching applied to the Sinkhorn update as a cool detail.
More from Models
- DeepSeek V4 training details: RL beyond collapse, WSD schedule, 1M context over 10T tokens — stochasticchasm · 2026-09-11
- Neel Nanda replicates Astra system card: no-CoT reasoning jumps 1.75x over next-best models — NeelNanda5 · 2026-09-11
- OpenAI pauses Pro subscriptions due to overwhelming demand — Charuru · 2026-09-11
- Artificial Analysis isn't broken: self-funded benchmarks, $13k spent on one model — Antblue · 2026-09-11
- Claude checkpoint diagnostician jokes Opus 3 is 'the cure' after newer model quirks — repligate · 2026-09-11
- Hume's Voice Replication Leaderboard: most natural clone ranked 8th of 11 on identity — realmrfakename · 2026-09-11