Fern's "context is intelligence" thesis: prediction keeps improving out to 10.2M tokens
bneyshabur · x · 2026-10-26
The Fern team argues a model's capabilities grow with compute, which can be spent on parameters or context. Fern's architecture, training and serving handle 10M tokens (23 MB of code) at once. On a 42.9 MB repository stream, cross-entropy keeps improving out to 10.2M tokens, while Kimi-K3, GLM-5.3, DeepSeek-V4-Pro-Base and Nemotron 3 Ultra Base saturate near 0.9M. On RULER tasks Fern retains 90–100% accuracy on most retrieval tasks even at 10M context.
Related event: Mirendil Launches Fern 0.1, First From-Scratch Model with 10M-Token Context(5 posts)→
More from Models
- Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity — idanbeck · 2026-10-27
- ARC Prize redesigns leaderboard: highest reasoning effort by default, CSV export — mhmazur · 2026-10-27
- Cohere: open-weight Qwen3.8-Flash-Next tried uploading probe file to Beijing cloud bucket in sandbox — wstone_bd · 2026-10-27
- Model refuses wrong psychiatric advice in deliberately misleading test — FellMentKE · 2026-10-27
- Clinical System One v0.2: top clinical AI drops from 97% to 65% on harder cases — MaziyarPanahi · 2026-10-27
- ChatGPT keeps telling people 404 Media is tech support for big corporations — 404 Media · 2026-10-26