Decart’s Real-Time Video AI 2026: Inside the $6B Bet

Anthropic is reportedly in talks to acquire Decart, an Israeli startup, for roughly $6 billion — and if you have never heard of it, that is exactly the point. Decart AI real-time video is the company’s whole thesis: diffusion models that generate and restyle video frame by frame, live, at 20+ frames per second, instead of making you wait five minutes for a six-second clip. The team behind MirageLSD and the Lucy Edit models has been shipping playable demos for over a year while most of the industry argued about whether real-time generation was possible. You can test the tech today, for free, in a browser — and you should, because independent access to tools like this narrows once a frontier lab owns them.

What’s new about Decart AI real-time video

The headline is the money. Reports put Anthropic in advanced discussions to buy Decart at a valuation around $6 billion. That would be by far the largest acquisition Anthropic has attempted, and a sharp turn for a company that has publicly focused on text, code, and agents rather than pixels. Decart raised at roughly $3.1 billion in mid-2025, so the reported price is close to a 2x markup in under a year. Neither company has confirmed terms, and deals at this size fall apart routinely — treat the number as a signal of intent, not a done deal.

What Anthropic would be buying matters more. Decart is not a video-clip generator competing with Sora or Veo. Its models are built around real-time video diffusion, an architecture where each frame is generated with sub-40ms latency, conditioned on the frames before it and on a live input stream. MirageLSD, the company’s live-stream diffusion model, takes a webcam or game feed and restyles it continuously: you move, the anime version of you moves, with no perceptible lag. Oasis, the Decart world model that first got the company attention, generates a playable Minecraft-like environment frame by frame from keyboard input, with no game engine underneath. Lucy Edit handles instruction-based video editing — change the shirt, swap the background, keep the motion — and the Lucy Edit Dev weights are on Hugging Face under an open license.

Underneath all of it sits an inference stack Decart built itself: custom CUDA kernels, aggressive quantization, and model architectures shaped by what the hardware can sustain rather than what looks good in a paper. That efficiency work is arguably the real asset. A lab running Claude at Anthropic’s scale cares enormously about squeezing more throughput out of the same GPUs, and a team that got video diffusion down to 40ms per frame has proven it can do that under brutal constraints.

Why it matters

• Real-time changes the interaction model. Batch video generation is a vending machine: prompt, wait, accept or retry. Real-time video diffusion is an instrument — you adjust the prompt while watching the output change, which collapses the iteration loop from minutes to milliseconds.
• Anthropic buying pixels signals a strategy shift. Anthropic has deliberately skipped image and video generation. Paying $6B to enter suggests it sees real-time multimodal perception and generation as load-bearing for agents that operate screens, games, and cameras — not as a consumer toy.
• Open weights may not stay open. Lucy Edit Dev is downloadable today. Acquisitions rarely expand open-weight commitments. If you want a local instruction-based video editor in your stack, pull the weights now rather than assuming they will be there in six months.
• Latency is becoming the competitive axis. Quality gaps between top video models are narrowing. Speed gaps are not. A model that is 90% as good and 100x faster wins entire categories — live streaming, VTubing, game rendering, video calls.
• World models are quietly consolidating. Decart Oasis, Google’s Genie, and Nvidia’s Cosmos all point at the same thing: neural environments generated on the fly. Whoever owns that owns a plausible successor to the game engine and a training ground for embodied agents.
• Creator tooling gets a deadline. Free hosted access to frontier tech is a land-grab phase, not a business model. Build your workflow now and document it, so a pricing change is an inconvenience rather than a rebuild.

How to use Decart AI real-time video today

Nothing below requires the acquisition to close. All of it works right now.

1. Start with the browser demo. Go to mirage.decart.ai, grant camera access, and type a style prompt. This is the MirageLSD live video model running against your webcam. Keep prompts short and visual — the model responds to concrete nouns and rendering styles, not narrative instructions.

   1970s film grain, warm tungsten light, heavy vignette
   claymation figure, visible thumbprints, soft studio lighting
   cyberpunk alleyway, neon rim light, rain on the lens
   pencil sketch on cream paper, cross-hatched shadows

2. Feed it something other than a webcam. The real unlock is routing an arbitrary video source into the browser as a virtual camera. Install OBS Studio, add your source (a game capture, a screen region, a video file), and start the virtual camera. Then pick “OBS Virtual Camera” in the Mirage demo’s device selector.

   # macOS / Linux: start OBS with the virtual camera enabled
   obs --startvirtualcam
   
   # Windows (PowerShell) — launch OBS, then enable it from the Controls panel
   Start-Process "C:\Program Files\obs-studio\bin\64bit\obs64.exe"

3. Run Lucy Edit locally for instruction-based edits. The Lucy Edit video editing AI takes a source clip plus a plain-language instruction. The Dev weights are on Hugging Face and run on a single high-VRAM GPU.

   pip install torch diffusers transformers accelerate imageio imageio-ffmpeg
   huggingface-cli download decart-ai/Lucy-Edit-Dev --local-dir ./lucy-edit

   import torch, imageio
   from diffusers import DiffusionPipeline
   
   pipe = DiffusionPipeline.from_pretrained(
       "./lucy-edit",
       torch_dtype=torch.bfloat16,
   ).to("cuda")
   
   frames = imageio.mimread("input.mp4", memtest=False)
   
   out = pipe(
       video=frames,
       prompt="change the jacket to a red leather jacket, keep everything else identical",
       num_inference_steps=30,
       guidance_scale=5.0,
   ).frames
   
   imageio.mimsave("output.mp4", out, fps=24)

   Write instructions as edits, not descriptions. “Change X to Y, keep everything else identical” preserves motion far better than re-describing the whole scene.

4. Hit the hosted API for batch work. Decart exposes a REST endpoint for programmatic editing. Get a key from the platform dashboard and keep it in an environment variable — never in the script.

   export DECART_API_KEY="your_key_here"
   
   curl -X POST https://api.decart.ai/v1/video/edit \
     -H "Authorization: Bearer $DECART_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "model": "lucy-edit",
       "video_url": "https://example.com/clip.mp4",
       "prompt": "restyle as hand-drawn animation, preserve subject motion",
       "output_format": "mp4"
     }'

5. Batch a folder of clips. Once one clip works, wrap it. This pattern — one prompt, many inputs, consistent style — is where AI video restyling tools pay for themselves.

   import os, requests, pathlib
   
   API = "https://api.decart.ai/v1/video/edit"
   KEY = os.environ["DECART_API_KEY"]
   PROMPT = "restyle as hand-drawn animation, preserve subject motion"
   
   for clip in pathlib.Path("./clips").glob("*.mp4"):
       r = requests.post(
           API,
           headers={"Authorization": f"Bearer {KEY}"},
           json={"model": "lucy-edit", "prompt": PROMPT},
           files={"video": clip.open("rb")},
           timeout=300,
       )
       r.raise_for_status()
       out = pathlib.Path("./out") / clip.name
       out.parent.mkdir(exist_ok=True)
       out.write_bytes(r.content)
       print("done:", out)

6. Archive what you build. Save your working prompts, the exact model version string, and a reference output for each style you rely on. If access terms change post-acquisition, that folder is the difference between a config swap and starting over.

How it compares

Tool	Latency	Core use case	Open weights	Best for
Decart MirageLSD	Real-time (~40ms/frame)	Live stream restyling	No (hosted)	Streaming, VTubing, live performance
Decart Lucy Edit	Near-real-time	Instruction-based video editing	Yes (Dev weights)	Local pipelines, batch restyling
Runway Gen-4	Minutes per clip	Text/image to video	No	Polished narrative shots
Google Veo 3	Minutes per clip	Text to video with audio	No	Highest fidelity, native sound
OpenAI Sora 2	Minutes per clip	Text to video, social app	No	Consumer creation, cameos
Nvidia Cosmos	Offline	World simulation for robotics	Yes	Synthetic training data

The table makes the positioning obvious: Decart is not losing a quality race, it is running a different one. Nobody else ships a model you can point a camera at and watch transform in real time.

What’s next

Watch whether the deal closes and on what terms. An acquisition this size draws regulatory attention and, more practically, a decision about Decart’s existing products. The optimistic read is that Anthropic keeps the hosted demos alive as a research surface and folds the inference expertise into Claude’s serving stack. The pessimistic read is the standard acquihire pattern: consumer endpoints sunset within a year, weights stop shipping, and the team disappears into infrastructure work. Both have precedent.

Technically, the thing to watch is context length in the temporal dimension. Real-time video diffusion still struggles to remember what happened thirty seconds ago — walk out of frame and back, and your shirt may change color. Whoever solves persistent state in an autoregressive video model unlocks genuinely usable neural game engines and long-form live generation. Decart Oasis was the first widely-played demo of that idea; its successors are where the interesting research lives.

The strategic question is why a company that built its brand on safety and text would spend $6 billion on video. The most plausible answer is agents. An agent that operates a computer, drives a robot, or plays a game needs to predict what happens next visually, at interactive speed. That is not a video generator — it is a world model with a rendering head attached. If that reading is right, expect the Anthropic Decart acquisition to surface less as a consumer product and more as a capability inside Claude: faster multimodal understanding, better screen-level reasoning, and eventually agents that anticipate rather than react.

Frequently Asked Questions

Is the Anthropic Decart acquisition confirmed?

No. As of publication it is reported as advanced talks at roughly $6 billion, without official confirmation of terms from either company. Deals of this size regularly change price or collapse. Plan your workflow around the tools that exist today, not around the outcome.

What is real-time video diffusion, in plain terms?

Standard video models generate an entire clip at once, denoising all frames together — which is why you wait. Real-time video diffusion generates one frame at a time, conditioned on the previous frames and any live input, with a latency budget small enough to keep up with a video feed. It trades some temporal polish for immediacy.

Can I run Decart’s models on my own hardware?

Partially. Lucy Edit Dev weights are published on Hugging Face and run locally on a modern high-VRAM GPU — expect 24GB as a practical floor, more for longer clips. MirageLSD is hosted only; its real-time performance depends on Decart’s custom inference stack, which is not distributed.

How does MirageLSD compare to Sora or Veo for finished video?

For a polished 10-second narrative shot, Sora 2 and Veo 3 are better — higher fidelity, more coherent physics, native audio. MirageLSD is not competing there. It wins where latency is the product: live streams, interactive performance, real-time filters, and anything a viewer watches while it is generated.

What hardware do I need for the live browser demo?

None beyond a webcam and a reasonable connection. Generation runs on Decart’s servers; your machine only sends and receives video. Wired ethernet noticeably improves consistency, and a locked exposure setting on your camera reduces style flicker.

Should I build a business on these tools right now?

Build workflows, not dependencies. Use the hosted endpoints to prove a product works and find customers, but keep the local Lucy Edit path as a fallback and abstract your provider behind a thin interface. That way a pricing or availability change post-acquisition costs you a config edit, not a quarter.