Gemini 4 Neon takes #2 on HearingBench
Google's mid-tier reasoning model reached the second spot on Adam Holter's HearingBench leaderboard, trailing only Gemini 4 Carbon. The benchmark challenges vision-language models to identify sounds, words, and full sentences purely from visual spectrogram images without audio input.
Scoring near the top of HearingBench proves Gemini 4 Neon's visual reasoning transfers remarkably well to acoustic signal processing without dedicated audio tokens.
- –HearingBench evaluates 100 audio clips across five difficulty tiers—from ambient noises to spoken sentences—strictly represented as spectrogram images.
- –Neon trails only its flagship sibling Gemini 4 Carbon, demonstrating that Google's mid-tier architecture retains top-tier visual and spatial pattern recognition.
- –High spectrogram literacy unlocks practical multimodal workflows in acoustic anomaly detection, audio forensics, and bioacoustics without requiring native audio encoders.
- –The evaluation highlights Gemini 4's structural advantage in cross-modal visual reasoning over competitors that struggle to parse dense frequency time-series visually.
DISCOVERED
5h ago
2026-11-19
PUBLISHED
6h ago
2026-11-19
RELEVANCE
AUTHOR
AdamHoltererer