No. 05 · Vision-language model
Search a 30-day archive in plain language.
LENS LM is the vision-language substrate beneath the rest of the stack. "Show me every red truck that turned left between 2pm and 4pm last Tuesday." Answered, with frames and reasoning, in seconds.
Capabilities
The substrate, productised.
01
Plain-language query
Ask in English, Hindi, Arabic, Spanish — the model speaks the operator's language.
02
Cross-camera search
Re-identify a person, vehicle or object across thousands of feeds.
03
Temporal reasoning
"Before / after / between" understood in archive context.
04
Evidence return
Every answer is grounded in clipped frames, with timestamps and confidence.
05
On-prem inference
The model runs in your DC. Queries don't leave the perimeter.
06
Substrate API
Powers VIEW, ICCC and SAFER reasoning surfaces internally.
Specs
What it can do.
Modalities
Visible, IR, thermal, audio · image and video reasoning together
Archive depth
Default 30 days, scalable to 12 months on cold storage with re-index
Query latency
<3s on a 7-day window across 200 cameras · A100/L40 class GPU
Languages
English, Hindi, Arabic, Spanish, French, German, Tamil · expanding
Deployment
On-prem first, sovereign cloud, hybrid · air-gapped supported
More on the stack