Elorian, an AI startup, has launched its first model, Glimpse, which is designed to reason about images by continuously referencing the source material rather than converting it to tokens. This approach allows Glimpse to perform actions like cropping into specific regions or analyzing video frames mid-reasoning, a capability Elorian calls a "visual scratch space." The company claims Glimpse is smaller and cheaper to train than frontier models and outperforms systems like GPT-6 Astra on visual reasoning tasks, aiming to eventually enable AI to not just see but also act in the real world. AI
IMPACT This model's novel approach to visual reasoning could set new benchmarks for multimodal AI capabilities.
RANK_REASON First model release from a startup founded by a former Google DeepMind researcher, with significant seed funding. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- Altimeter Capital
- Andrew Dai
- Elorian
- Glimpse
- Google DeepMind
- GPT-6 Astra
- Haiku
- Jeff Dean
- Luna
- Menlo Ventures
- Nvidia
- sonnet
- Striker Venture Partners
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →