Chandra 2.5 cuts OCR script confusions 7x, beating Gemini 3.8 Flash and Opus 5.5
VikParuchuri · x · 2026-10-16
Datalab shipped Chandra OCR 2.5 with a novel training technique that fixes "script confusions"—when an uncertain VLM picks a token that's plausible to it but looks nothing like the page, e.g. outputting Chinese for an English word. It's live on their API.
On a hard set of 5,600 real document pages, Chandra 2.5 has a 0.16% script-confusion rate—7x fewer than Chandra 2.4 (1.19%)—and below Gemini 3.8 Flash (0.61%) and Claude Opus 5.5 (0.89%). LightOnOCR-3 scored 3.04% and DeepSeek-OCR 2 6.08%, with confusions often spanning whole passages. Manual verification confirmed every Chandra flag was real, while competitors' flags were largely false positives.
More from Models
- NEURA launches Neuron AI, a multimodal PLA model giving robots touch, force feedback and memory — CyberRobooo · 2026-10-17
- Sentience claims per-user models: train, continually update and serve one model per person — soleio · 2026-10-17
- Four Gemini 4 Argon agents beat Minecraft in 27 hours, writing 2,783 programs — GoogleDeepMind · 2026-10-17
- Meta open-sources Muse Spark Open to mixed reception: 'trash' license, nerfed benchmark scores — gabriberton · 2026-10-17
- If new diffusion models are open-weight only, what's left for commercial use? — blackpanda8484 · 2026-10-17
- Reflection Beam OSS architecture: sliding window attention, gated attention, QK Norm — A_K_Nain · 2026-10-17