Elorian Glimpse

Visual thinking that stays grounded in images and video

Try now

Read Docs

About model

Glimpse is a research preview from Elorian, which describes it as a new category of visual thinking model: a system that reasons about the physical world from what it sees. It follows objects in motion and across changing viewpoints, understands how parts fit together, and connects an observation to what happens next. Rather than reasoning only in text, Glimpse brings visual representations into the reasoning process itself, revisiting and refining them as its thinking unfolds: it looks closely, questions its first impression, and returns to the visual evidence before it answers. As an agent, it is trained end to end to use multimodal tools that write code, transform images, and watch videos, and it can coordinate subagents that exchange text, visual artifacts, and internal representations. Post-trained on an open-weights base model. Available on Together AI.

‍

Visual Reasoning

Native

Refines internal visual representations as it thinks, not just text

Temporal Understanding

Video

Tracks identity through motion, camera moves, and changing viewpoints

Subagent Coordination

Multi-Agent

Subagents share text, visual artifacts, and internal representations

Model key capabilities

Grounded Visual Reasoning: Reasons in visual space, zooming into regions and returning to the visual evidence to check its interpretation before answering
Motion & Physical Understanding: Follows objects through motion and viewpoint changes, and connects an observation to its likely physical consequence
Multimodal Agent Tools: Trained end to end to use tools that write code, transform images, and watch videos, and to coordinate subagents on a shared problem
Available on Together AI: Accessible through dedicated deployment on Together AI

Model card
Architecture Overview:
• Visual thinking model post-trained on an open-weights base model
• Brings visual representations into the reasoning process, revisiting and refining them as thinking unfolds
• Native image and video input with text output; inspects video selectively, rewatching short intervals at denser sampling and tracking single features across time

Training Methodology:
• Trained end to end to use multimodal tools that write code, transform images, and watch videos
• Trained to spawn and coordinate subagents that exchange text, visual artifacts, and internal representations
• Focused on grounded reasoning about objects and their spatial relationships

Performance Characteristics:
• All figures Elorian-reported at maximum thinking effort
• Robotics (WGO Bench end-to-end manipulation events: 42.70; ERQA embodied spatial reasoning: 83.75)
• Visual perception (BabyVision fine-grained visual reasoning: 95.36; HSS scene understanding across image and video: 46.36)
• Scientific reasoning (CharXiv-R scientific chart reasoning: 92.90)
• Video captioning (UGC-VideoCap short-form: 82.66; DREAM-1K dense captioning: 76.04)
• Temporal reasoning (TOMATO fine-grained temporal reasoning: 93.60; Video-MME v2 with subtitles: 56.68; E.T. Bench event understanding: 63.55)
• Counting (FSC-147 open-vocabulary counting: 83.36; CountQA product image counting: 82.66)
‍
Prompting
Together AI API Access:
• Access Glimpse via Together AI APIs; endpoint details are provided with your deployment
• Authenticate using your Together AI API key in request headers
• Supports text, image, and video input with text output
• Available on Together AI through dedicated deployment
‍
Applications & use cases
Robotics Data & Planning:
• Label and segment teleoperation data, naming objects and splitting episodes into timestamped pick-and-place events
• Act as a high-level planner for vision-language-action policies, breaking requests into steps and checking when each is done
• Identify completed actions in robot footage

Video Understanding at Scale:
• Generate dense captions covering fine-grained movements and social or emotional dynamics
• Tag short-form and social video for search and organization
• Track and count objects across handheld or moving-camera footage

Scientific & Operational Inspection:
• Analyze scientific charts and laboratory data
• Count equipment and inventory precisely, separating lookalikes and object states
• Reason about physical consequences from photographs and clips
‍