Reka AI Unveils Video Reasoning: MotoGP Rider ID Accuracy Hits 90.9%
RekaAILabs · x · 2026-07-29
Reka AI published a blog post detailing its video models' capabilities in spatial grounding and temporal reasoning, arguing that the bottleneck in video AI has shifted from recognition to reasoning.
- Spatial Grounding: The model supports open-vocabulary instructions (e.g., "Detect: red car") and outputs precise bounding boxes and pixel-accurate segmentation masks. On the RefCOCO-A benchmark, its lightweight Reka Edge model scored 93.01, beating Gemini 3 Pro's 81.46.
- Temporal Reasoning: The model can track and understand how events evolve over time within the footage.
- Real-World Impact: In MotoGP broadcasts, by combining 4 core reasoning signals (space, time, identity, and judgment), the model boosted rider ID accuracy from 39.6% to 90.9% without increasing model size.
- Edge Deployment: These reasoning capabilities are specifically designed to run on the edge, directly on the devices where the cameras are located.
More from Models
- GitHub Copilot Integrates Grok 4.5: Fast, Smart, and Affordable — intellectronica · 2026-07-29
- Taming Anthropic's Opus 5: Brilliant in Flashes, Frustrating in Practice — every · 2026-07-29
- User reverses course and says GPT 5.6 Sol is actually a really good model — TheZachMueller · 2026-07-29
- Sam Altman teases GPT-5.6 Sol on Cerebras at 750 tokens/sec in July — daniel_mac8 · 2026-07-29
- Grok 4.5 Medium tops LaurenBench with 56.9%, ahead of Claude Sonnet 5 and GLM 5.2 — elonmusk · 2026-07-29
- A viral chart compares 12 paid AI tools with free replacements — nikola_mr64990 · 2026-07-29