Read once, decide instantly
Core model parameters ~9B
Model files 6.50 GB
Turn text or images into structured decisions; process several questions together and compare candidate answers with option probabilities
- Text & images
- Multiple questions
- Structured answers
Recorded on the board
Screen recording of the board web demo, recorded on Qualcomm QCS8550.
Highlights
Text and image input
Use background text on its own or add a JPEG/PNG image for scene understanding, object checks and diagram questions.
Multiple questions together
Process several questions in one SDK request, sharing the same background and image.
Three decision types
Choose among named candidates, answer yes/no questions or score ordered levels, with structured results for your application.
Visible option probabilities
Compare probabilities for supplied candidates and highlight the most likely answer, without generating open-ended text.
On-device inference
Run locally on Qualcomm chips without a cloud inference service.
From demo to integration
Four templates cover scenes, objects, diagrams and a support ticket, alongside a C API and Python binding.
Pipeline
From input to output: where every step runs.
- 01Browser
Input and questions
Enter background text, an optional image and candidate answers
- 02CPU
Preprocessing
Decode images, tokenize text and assemble the request
- 03NPU
Text and image encoding
Run the vision encoder and language graphs on the device
- 04CPU
Candidate scoring
Compute option probabilities and assemble choices, yes/no answers or scores
- 05Browser
Results
Display structured answers, option probabilities and inference timings
Supported hardware
- Core model parameters
- ~9B
- Model files
- 6.50 GB
- License
- Pandora License