LFM3-VL-2B scores just 7% non-hallucination rate, confusing a deadly mushroom for food
crusaderky · reddit · 2026-11-18
A Reddit user flagged that Liquid AI's LFM3-VL-2B (a phone-runnable vision model) scores an alarming 7% non-hallucination rate (1 - hallucination rate) on Artificial Analysis's AA-Omniscience v1.1 benchmark — dead last in the comparison chart, where LFM3-1.7B hits 92%, Nanbeige4.3-33B 82%, and Qwen3.8 27B 71%.
Real-world failures cited: identifying the deadly amanita phalloides as a tasty mushroom recommended for risotto, and reading a "trespassers will be shot" sign as "no photos beyond this point." The poster asks for first-hand confirmation of the scores.
More from Models
- Asked Opus to make a video explaining consciousness and existence — "no wiser but cool" — flowersslop · 2026-11-18
- theo retests GPT-6.2 Astra: better, but still 'directionally very bad' at debugging — airesearch12 · 2026-11-18
- Claude correctly folds ubiquitin from a plain prompt, no specialized tools needed — iScienceLuvr · 2026-11-18
- GPT-6.2 Astra Tops Bug-Hunt Coding Benchmark at 62.7/105, One-Fifth Sonnet's Cost — PawelHuryn · 2026-11-18
- OpenAI's enterprise Dots reportedly to cost $50/$200/$500 per tier plus usage — and businesses are turning them off — jdjohnson · 2026-11-18
- Flint 40B: Reddit users flag a coding-focused MoE model rivaling Qwen 3.6 35B locally — soyuz-1 · 2026-11-18