LeVJEPA 1.5 Large matches V-JEPA with 5.6x less compute, sparking efficiency debate
randall_balestr · x · 2026-11-03
Lukas Kuhn released LeVJEPA 1.5 Large, closing the gap to V-JEPA on image and video benchmarks using 5.6x less compute and over an order of magnitude less data; checkpoint is available. Reposter argues this shows the field should focus on improving pretraining losses rather than just adding data and GPUs, calling SIGReg a good first step.
More from Research
- Qwenputer compiles a full RISC-V CPU into Qwen3.5-4B weights, bit-exact, even running another LLM — udmrzn · 2026-11-03
- Astra IDs Breakout, Pong, Pac-Man and Freeway from raw model weights alone — burny_tech · 2026-11-03
- World model that predicts the future directly in 3D, fully open-sourced — YunzhuLiYZ · 2026-11-03
- Scalable gradient attribution finds deceptively harmful training data, with a 0.6B model as monitor — PeterHndrsn · 2026-11-03
- Dynamical releases SDL-2: frontier agents make correct materials-science calls in only ~half of experiments — BenBlaiszik · 2026-11-03
- Follow-up: Astra's weight edits only worked on Qwen's visible flaw, not on Gemma 270m — teortaxesTex · 2026-11-03