Jev reportedly trained on 100% synthetic data with no pretraining, just post-training
MaziyarPanahi · x · 2026-09-20
Replying to Sebastian Raschka, MaziyarPanahi notes Jev's generalization is impressive: the founder hinted its dataset was 100% synthetic, and the model reportedly skipped pretraining entirely, doing only post-training on an existing model — a route praised as a huge cost saver. Details remain community-sourced and unconfirmed.
Related event: Raschka would spend a $100M LLM budget entirely on post-training(3 posts)→
More from Infra
- Turbovec: Rust vector index fits 10M docs in 4GB and beats FAISS by 3.4x at 4-bit — bibryam · 2026-09-21
- Google's Agent Substrate detailed: AX app layer on managed agentic compute infra — rakyll · 2026-09-21
- Why sandbox-as-a-service startups are booming — and whether labs will just build it themselves — dejavucoder · 2026-09-21
- Agents may discover million-times-cheaper training, making data centers look silly, predicts Steve Moraco — menhguin · 2026-09-21
- llama.cpp PR enables sparse FlashAttention for Qwen, another inference speedup — jacek2023 · 2026-09-21
- Open-source Jev model runs offline on Mac M4 via CoreML at 45 decisions/sec — putna · 2026-09-20