Glimpse: A Visual Reasoning Model That Uses a Visual Scratch Space
Melvin Vivas · X video post · 2026-10-16 · 1:29 · 395 views · Open on X
Topics: Industry Trends & Job Market, LLM Fundamentals · Level: intermediate
Summary
Melvin Vivas shares a launch video for Glimpse, the first visual reasoning ("visual thinking") model from Illurion AI. It is presented by co-founder and CEO Andrew Dye, who previously spent 12 years at Google Brain and Google DeepMind. The video explains how the model reasons about the physical world: it uses a "visual scratch space" that feeds crops, selected frames and new views back into its reasoning. It also gives claimed benchmark, cost and compute numbers.
Key points
- Glimpse is pitched as the first in a new category of vision-specialized models that reason about the physical world from what they see.
- Abilities these models need: following objects as they move and as the viewpoint changes, understanding how parts fit together, and linking an observation to what happens next.
- Core technique: a 'visual scratch space'. Reasoning steps are conditioned on transformed images (crops, selected frames, new views) that are brought back into the reasoning loop.
- Claimed results: frontier-level performance, competitive with 'Astra', on benchmarks covering counting, scientific chart reasoning, and video and robotic captioning.
- Claimed cost: about 1/16 of Astra's token cost.
- Built by a small team using only a few hundred GPUs.
- Intended uses: robotics, science, engineering, and other work that needs physical-world understanding.
- All performance numbers are the vendor's own launch claims and have not been checked independently.
Resources mentioned
- Glimpse · tool · elorian.ai · check price
Illurion AI's visual reasoning model, which reasons over images and video using a visual scratch space. - Illurion AI · website · elorian.ai · free
The AI startup that built the Glimpse visual reasoning model (company name as spelled in the machine transcript). - Andrew Dye · person · dblp.org · free
Co-founder and CEO of Illurion AI, formerly at Google Brain and Google DeepMind for 12 years (name as transcribed). - GPT-6 Astra · tool · openai.com · paid
The model announced in the quoted launch post, pitched as the developer's most capable model for work, coding, science and cybersecurity, and able to operate a computer.
Also in: Is AI Getting Cheaper? How to Measure LLM Cost Per Task and Per Intelligence Level (Melvin Vivas on X · notes), GitHub Copilot Adds Intelligent Local Model Routing (Project HydraFusion) (Melvin Vivas on X · notes), Comparing Opus 5.5, GPT-6 Astra and Grok 4.7 on OpenRouter (Melvin Vivas on X · notes), OpenAI Makes GPT-6 Astra and GPT-6.1 Sol About 50% Faster (Melvin Vivas on X · notes) and 55 more
More in Industry Trends & Job Market
- DeepSeek-V4.1-Flash-1017 release: better benchmarks, shorter outputs
- Gemini 4 Argon now available in Devin
- Reflection AI's Beam: a new 501B MoE open-weights coding model
- Grok Bot vs OpenAI in the consumer AI agent market
- Why Simpler Agent Bots Appeal to Non-Technical Users
- Why Grok Bot Could Be the Next "ChatGPT Moment" for Personal AI Assistants