Decart
Last reviewed
May 16, 2026
Sources
25 citations
Review status
Source-backed
Revision
v1 ยท 3,499 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 16, 2026
Sources
25 citations
Review status
Source-backed
Revision
v1 ยท 3,499 words
Add missing citations, update stale details, or suggest a clearer explanation.
Decart (also known as Decart AI, operating from the domain decart.ai) is an Israeli artificial intelligence company headquartered in Tel Aviv that builds real-time generative AI systems for video and interactive worlds. Founded in late 2023 by Dean Leitersdorf and Moshe Shalev, two veterans of Israel's signals intelligence Unit 8200, Decart became known in 2024 for Oasis, a real-time, fully AI-generated playable open world inspired by Minecraft that the company released with chip startup Etched. It has since shipped a sequence of real-time video models, including MirageLSD (Live Stream Diffusion) and the Lucy family, that transform live video feeds at broadcast quality with sub-100 millisecond latency.
By August 2025, Decart had closed a $100 million Series B led by Sequoia Capital at a reported $3.1 billion valuation, less than two years after exiting stealth. The company positions itself as a real-time alternative to clip-by-clip generative video tools like Runway, Luma Dream Machine, Pika 2.5, and the avatar-focused Hedra Character platform, and frames its work as an early step toward general purpose world models.
Decart's central technical claim is that it can run high quality diffusion models for video at the frame rate of a normal video feed, rather than the multi-second per clip latencies typical of offline generators. The company combines its own model architectures with a proprietary GPU optimization stack that it also licenses to third parties. According to coverage by Calcalist and Fortune, this licensing business produced revenue from the company's first year of operation through multi-million dollar contracts with cloud providers and AI labs.
Decart's public products fall into three groups: interactive world demos such as Oasis, where the entire frame is generated by a neural network in response to user input; real-time video transformation models such as MirageLSD and Lucy 2, which restyle an incoming video stream from a camera, screen capture, or game; and a developer platform (Mirage and the Decart API) that exposes these models to third party developers and partners such as Crusoe Cloud.
Decart was incorporated in late 2023 in Tel Aviv. The two co-founders are Dean Leitersdorf, who serves as chief executive officer, and Moshe Shalev. Both came out of Israel Defense Forces Unit 8200, the country's signals intelligence and offensive cyber unit that has long been a feeder for Israeli technology startups. Leitersdorf, who was in his mid-twenties at founding, holds three computer science degrees including a PhD completed at age 23 at the Technion. Profiles in Calcalist and Ynet describe a deliberately spartan operating culture, with most engineers working out of a single Tel Aviv office on long, on-call schedules.
Decart emerged from stealth on August 6, 2024, simultaneously announcing a $21 million seed round led by Sequoia Capital with participation from Zeev Ventures and several angels. At the time the company said it would focus on "making AI cheap enough to be available to everyone" through a combination of model and systems work, citing very large cost reductions in serving inference for large foundation models relative to standard implementations.
On October 31, 2024, Decart and Etched released Oasis, described by the two companies as the first playable, real-time, open-world AI model. Oasis generates each frame of a Minecraft-like game on the fly using a transformer-based diffusion backbone, with no underlying game engine: physics, rendering, lighting, world layout, inventory, and game logic are all produced by the model in response to keyboard and mouse input from the player. The architecture combines a Vision Transformer based spatial autoencoder with a Diffusion Transformer backbone, trained on millions of hours of Minecraft footage paired with input events.
At launch, Oasis ran at roughly 20 frames per second at 360p on an NVIDIA H100 GPU, with inference time per frame around 47 milliseconds. The companies positioned the model as a proof point for Etched's forthcoming Sohu transformer-only ASIC, which they said could serve substantially larger Oasis variants in real time at higher resolutions. Decart released code and a 500 million parameter checkpoint for local use along with a hosted demo of a larger model. The demo went viral: Tom's Guide called it "the future of gaming," TechSpot described it as a "nightmarish hallucination" because of long-horizon drift, and TechCrunch reported that the hosted demo reached one million users in three days.
On December 19, 2024, TechCrunch reported that Decart had raised a $32 million round led by Benchmark at a $500 million post-money valuation, less than two months after closing its seed. The article framed the round as funding for "AI tech and open world apps," with Decart telling reporters that it planned to broaden Oasis-style interactive worlds and to ship real-time video products beyond gaming. After this round Decart had raised approximately $53 million in total.
In 2025, Decart introduced MirageLSD, which it described as the first Live Stream Diffusion model. MirageLSD takes any incoming video, from a webcam, a video chat, a screen share, or a video game, and transforms it in real time according to a text prompt, producing 24 frames per second output at roughly 40 milliseconds end-to-end latency per frame. The company published a technical note on a technique it calls history augmentation, in which the model is fine tuned on synthetically corrupted recent frames so that it learns to correct for drift and artifacts characteristic of autoregressive video generation. MirageLSD was made available through Decart's own platform at mirage.decart.ai and, via partnership, through Crusoe Cloud, which began offering the model to its high-performance compute customers as a hosted API.
On August 7, 2025, Decart announced a $100 million Series B led by Sequoia Capital, with participation from Benchmark, Zeev Ventures, and new investor Aleph. Fortune reported the post-money valuation at approximately $3.1 billion, more than six times the December 2024 mark, and put Decart's cumulative raise at about $153 million. SiliconANGLE and Globes both reported the round as a Series B focused on scaling Decart's real-time video platform.
Alongside MirageLSD, Decart introduced the Lucy line of real-time video transformation models. Lucy 2.0, announced in late 2025, generated and transformed live video at 30 frames per second at 1080p resolution with under 100 millisecond latency, according to Decart's own publication and reporting in The Decoder and Marketing Tech News. The model is built as a pure diffusion system without depth maps or 3D scaffolding, with motion and physics behaviors emerging from training on large amounts of video. Decart said sustained real-time generation with Lucy 2.0 ran at roughly three dollars per hour of generated video on its infrastructure. Also in 2025, Decart announced a multi-million dollar partnership with the Technion in Haifa, funding research, scholarships, and joint work on systems and generative video.
| Year | Product or model | Type | Notable specifications |
|---|---|---|---|
| 2024 | GPU optimization stack | Inference systems software | Licensed to cloud providers and AI labs |
| 2024 | Oasis (with Etched) | Real-time interactive world model | ViT autoencoder + DiT backbone; 20 fps at 360p on H100; ~47 ms per frame; 500M parameter checkpoint released |
| 2025 | MirageLSD | Live Stream Diffusion video model | Transforms any video stream from a prompt; 24 fps; under 40 ms latency per frame; history augmentation against drift |
| 2025 | Mirage / Decart API | Developer API for real-time video | Exposes Decart models; hosted via decart.ai and partner clouds |
| 2025 | Lucy 2.0 | Real-time video transformation model | 30 fps at 1080p; under 100 ms per frame; pure diffusion; ~$3 per hour of generated video |
Oasis is the product that put Decart on the map. The model accepts keyboard and mouse input from a user and produces, frame by frame, the next image that a Minecraft player would see. Unlike retrieval based systems or game engines wrapped with generative shaders, Oasis has no underlying simulation: each frame is the output of a denoising step conditioned on prior frames, the player's current action, and learned latents. A Vision Transformer based spatial autoencoder compresses frames into a latent grid, and a Diffusion Transformer predicts the next latent given a short window of history and the current input. Both components are pure transformer stacks, well suited to the transformer-only Sohu chip designed by Etched.
Decart and Etched reported inference times of about 47 milliseconds per frame on NVIDIA H100 GPUs, allowing for 20 frame per second playback at 360p, and said that on Sohu hardware, larger Oasis variants with more than 100 billion parameters could be served in real time at higher resolutions. Decart released the weights for a 500 million parameter Oasis model and the inference code under an open license, while keeping larger checkpoints behind the hosted demo. Because the model has no persistent state beyond a short context of frames, players who travel far in a generated world often see scenes lose coherence over time. TechSpot and others noted that this produces dreamlike or hallucinated landscapes after extended play, a known limitation of current world models that history augmentation is designed to mitigate.
MirageLSD is Decart's real-time video transformation model. It accepts a continuous video input, such as a webcam feed or game capture, and outputs a continuously transformed video stream that follows a user-supplied text prompt and optional reference image. The company describes the model as the first system to deliver infinite duration, real-time diffusion video at acceptable quality, with 24 frame per second output at end-to-end latency under 40 milliseconds per frame, frame-by-frame generation that preserves temporal coherence across long sessions via history augmentation training, and compatibility with arbitrary video sources including camera, screen, video chat, and game capture, with prompts that can change live. Decart positions MirageLSD as an enabling layer for live streaming, virtual production, interactive media, and what some early users have called "vibe-coded" gameplay, where a streamer can repaint the visual style of a game without modifying the underlying assets.
The Lucy family extends MirageLSD's real-time video work into a more general purpose model suitable for higher resolution and longer sessions. Lucy 2.0, the most recent generation as of this writing, runs at 30 frames per second at 1080p resolution, with sub-100 millisecond per-frame latency. Use cases described in Decart's launch material and coverage by Marketing Tech News, The Decoder, and AI World Today include live character swaps, virtual try-ons, product placement, environment swaps, and motion-preserving stylization of existing footage. Lucy 2.0 is implemented as a pure diffusion model. It does not rely on auxiliary depth maps, meshes, or explicit physics engines; instead, behaviors that look physical such as object permanence and cloth dynamics emerge from training on large quantities of video. Decart has reported a sustained generation cost of roughly $3 per hour of output, which it frames as more than an order of magnitude cheaper than earlier per-second pricing for generative video.
Decart exposes its models through an API platform branded Mirage at mirage.decart.ai. The platform hosts MirageLSD, Lucy, and image and video generation endpoints, and is the primary commercial surface for third-party developers. Decart has also partnered with Crusoe Cloud to make MirageLSD available within Crusoe's GPU cloud environment for enterprise customers. Underneath the models, Decart's earliest revenue source, as described in Calcalist and Fortune, is a proprietary stack for optimizing inference on large transformer models, licensed to cloud providers and AI labs to reduce the cost of serving large models. Decart has cited cost reductions for image and video generation work from tens or hundreds of dollars per hour of output down to under $0.25 per hour, though specifics of those benchmarks have not been independently verified.
As of August 2025, Decart had raised approximately $153 million across three disclosed rounds. The company is privately held and its detailed cap table has not been publicly released.
| Round | Date | Amount | Lead investor(s) | Other participants | Reported valuation |
|---|---|---|---|---|---|
| Seed | August 6, 2024 | $21 million | Sequoia Capital | Zeev Ventures, angels | Not disclosed |
| Series A (reported by TechCrunch as a follow-on) | December 19, 2024 | $32 million | Benchmark | Sequoia Capital, Zeev Ventures | $500 million post-money |
| Series B | August 7, 2025 | $100 million | Sequoia Capital | Benchmark, Zeev Ventures, Aleph | $3.1 billion post-money |
| Cumulative | August 2025 | ~$153 million | - | - | - |
Decart's most prominent investors are Sequoia Capital, Benchmark, Zeev Ventures, and Aleph. The company's Series B was reported by Fortune, SiliconANGLE, Globes, and The SaaS News among others.
| Name | Role | Background |
|---|---|---|
| Dean Leitersdorf | Co-founder and CEO | Israeli computer scientist; veteran of IDF Unit 8200; three computer science degrees including a PhD completed at age 23 at the Technion |
| Moshe Shalev | Co-founder | Veteran of IDF Unit 8200; co-launched Decart out of stealth in 2024 |
Decart's broader team is concentrated in Tel Aviv, with research and engineering staff drawn heavily from Israeli computer science programs and from Unit 8200 alumni networks.
Decart has signed a number of partnerships that combine its models with third-party hardware, cloud, or research infrastructure.
| Partner | Nature of partnership | Notes |
|---|---|---|
| Etched | Joint development and launch of Oasis | Decart's transformer-based Oasis model was co-launched with Etched and was used as a showcase application for Etched's Sohu transformer-only ASIC |
| Crusoe Cloud | Distribution of MirageLSD as a managed cloud service | Customers can subscribe to MirageLSD inference capacity through Crusoe's GPU cloud rather than directly through Decart |
| Technion (Israel Institute of Technology) | Multi-year research and education partnership | Decart pledged multi-million dollar funding to Technion in 2025 for systems and generative video research, scholarships, and joint work |
| Sequoia Capital, Benchmark, Zeev Ventures, Aleph | Equity investors | Lead and follow-on investors across seed, Series A, and Series B financings |
Decart competes in a busy generative video market but occupies an unusual position. Most well-known generative video products, including Runway's Gen series, Luma Dream Machine, Pika 2.5, and Google Veo, are designed to produce short clips offline from prompts or input images, with users waiting some seconds for a render. Hedra Character focuses on talking head avatars driven by audio. Decart's MirageLSD and Lucy models, by contrast, are aimed primarily at continuous, real-time transformation of a live video feed.
| Company | Core product | Primary mode | Approximate latency | Main use cases |
|---|---|---|---|---|
| Decart | MirageLSD, Lucy 2.0, Oasis | Real-time video and interactive worlds | Under 40 ms per frame (MirageLSD); under 100 ms at 30 fps 1080p (Lucy 2.0); ~47 ms at 20 fps 360p (Oasis) | Live streaming, real-time game restyling, interactive AI worlds |
| Runway | Gen-3 and Gen-4 video models | Offline clip generation, video-to-video editing | Seconds to minutes per clip | Brand and film advertising, VFX, professional editing |
| Luma Dream Machine | Ray and Ray3 family | Offline clip generation; HDR support in Ray3 | Tens of seconds to minutes per clip | Cinematic clips, marketing, social content |
| Pika 2.5 | Pika video model | Offline clip generation, stylized animations | Tens of seconds per clip | Short-form social content, meme animation |
| Hedra Character | Character-2 audio driven avatars | Audio-driven talking heads | Near real-time talking head playback | Talking avatars, dubbed video, presentations |
Industry coverage has framed Decart less as a head-to-head competitor with these tools and more as the leading example of a new sub-category, sometimes called real-time world models or live video AI. Articles in The Decoder, eWeek, and SiliconANGLE generally describe Decart's pitch as turning generative video from a render farm into something more like an instrument that can be played live.
Decart has been covered repeatedly by mainstream technology and Israeli business press, including TechCrunch's December 2024 article on the $32 million round, Fortune's August 2025 exclusive on the $100 million Series B at a $3.1 billion valuation, Tom's Guide and TechSpot reporting on Oasis (and its tendency to dissolve into surreal landscapes after long sessions), The Decoder's coverage of Lucy 2.0, and Calcalist profiles headlined "From stealth to $3.1 billion in less than a year." Ynet has described Decart as an Israeli AI unicorn aiming at large entertainment markets, comparing its ambitions to Netflix, YouTube, and TikTok.
Technical reception has focused on the difficulty of doing diffusion at frame rates and on long-horizon consistency. Engineers covering Oasis frequently note that the lack of an explicit world state means the model occasionally invents inventories, contradicts itself, or drifts into impossible terrain when a player walks far from their starting point. Decart has cited history augmentation and longer context windows as the path forward.
Decart's models share several recurring design choices. Both Oasis and the Lucy and Mirage video models are built on transformer-based diffusion backbones, often paired with transformer autoencoders, consistent with the company's bet on transformer specific accelerators such as Etched's Sohu. For autoregressive video generation, Decart fine-tunes models on synthetically corrupted recent frames so that the model learns to anticipate and correct artifacts in its own output, a technique it calls history augmentation that is central to MirageLSD's ability to run for long sessions without collapsing into noise. The company frames its GPU optimization stack as core intellectual property as important as the model weights themselves, and has released code and weights for smaller Oasis checkpoints while keeping flagship models hosted.
Decart frequently describes its long-term goal as building general purpose world models that can simulate arbitrary environments in response to user input, with current products such as Oasis, MirageLSD, and Lucy presented as steps toward that goal rather than endpoints. The company is headquartered in Tel Aviv and maintains close ties to Israeli universities, especially the Technion.