A simple float64 prime test that small local models keep failing
MrMrsPotts · reddit · 2026-10-17
A Reddit user shares a deceptively simple math test for small local models: "What is the largest prime that can be exactly represented in float64?" The question combines floating-point representation limits with prime reasoning, and many small models fail at the combination. The author invites others to report which local models still get it wrong.
More from Models
- Model 6.1 Sol Solves a 6D Rubik's Cube, Far Beyond Human Intuition — flowersslop · 2026-10-17
- 10 open-source repos for training small models to do your agent's job — victor_explore · 2026-10-17
- Claude Sonnet 4.5 will be fully removed from AWS Bedrock by April 2027 — repligate · 2026-10-17
- DeepSeek V4.1 Flash talks like a human while coding: "I love this", "Disaster", "Ugh" — teortaxesTex · 2026-10-17
- Theory-of-mind test: Claude spots AI-written text, ChatGPT and Gemini fail — Aizkmusic · 2026-10-17
- Rumor: Core automation to release a model literally called 'AGI' on Nov 10 — imjustnewatai · 2026-10-17