DigitalOcean Serverless tests Nemotron 3.5 Super VL on Grafana screenshots fail

An evaluation of 100 Grafana-style dashboard panels tested against four vision models on DigitalOcean Serverless Inference found that no model read key facts reliably enough to serve as an authoritative source of truth. The author suggests taking numerical values directly from alert payloads while using a fast, cost-effective vision model like Gemma 4 or Qwen3.8-Max purely to generate narrative descriptions, validated by cross-checking against alert data. Open article →