Research

How Much Can We Optimize?

At Decart, research is measured by what it makes possible. We advance the science of doing more with the compute already available. Every advancement feeds directly into what we build, making our models faster, more efficient, and more capable.

Publications

Coming soon
Research Soon
What if video editing had a volume knob?
Text-driven video editing lets you describe what to change. But what about how much? We introduce Adaptive-Origin Guidance (AdaOr), a method that turns edit intensity into a smooth, continuous slider.
Research Feb 18, 2026
Our goal

AI that perceives and reacts as fast as the world does.

Live is the standard every line of our research is solving for

Research pillars

01Lucy

Live Video

  • Object and background replacement
  • Consistency under motion and across frames
  • Virtual try-on that preserves logos fit and detail
  • Support for new input types and camera angles
  • Live VFX and selective editing
02Oasis

Physical AI

  • Short- and long-horizon memory
  • Physical accuracy in object interaction
  • Proprioception (spatial self-awareness)
  • Object interaction
03DOS / Cogito

Compute Efficiency

  • Higher throughput lower latency
  • Longer context support
  • Faster time-to-first-token
  • Higher output per chip
04Origin

Foundational Models

  • Improved text generation
  • Postures and images with specific styling
  • Models optimized for the DOS stack
  • Controllable image generation
“Live AI is not just a speed benchmark; it is a fundamental paradigm shift. When hardware and representation learning evolve together, AI ceases to be an offline tool and becomes a continuous, interactive reality.”
Chaim Baskin
Chaim Baskin
Decart Research | Academic Partnerships

The benchmark keeps moving.

Each number below was the ceiling - until it wasn't.

Frames per second
3 FPS → 101 FPS
Lucy running on AWS Trainium
Tokens per second
200 TPS → 1600 TPS
Frontier LLM inference on TPU
Latency
X sec → 40 ms
Lucy and Oasis running on GPU

Meet us there

Oct 20–22  ·  2026 NVIDIA GTC EMEA Nvidia’s upcoming premier regional event Research Berlin  |  Germany
Nov 10–12  ·  2026 CoRL Conference on Robot Learning Research Austin  |  Texas
Nov 30 – Dec 4  ·  2026 AWS re:Invent A top cloud conference Research Las Vegas  |  USA
Dec 6–12  ·  2026 NeurIPS World’s most prestigious annual academic conference for machine learning and computational neuroscience Research Sydney  |  Australia
Oct 2–8  ·  2027 ICCV International Conference on Computer Vision Research Hong Kong

Built by researchers, for researchersCommunity tools

Research tool

Unfold by Decart

Your research companion for any paper. Unfold gives you a quick preview, AI chat, citations, and an interactive graph - everything you need to understand a paper and explore the research around it, all in one place.

Open Unfold