Decart
Last reviewed
Jun 5, 2026
Sources
33 citations
Review status
Source-backed
Revision
v2 ยท 5,054 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 5, 2026
Sources
33 citations
Review status
Source-backed
Revision
v2 ยท 5,054 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. In May 2026 it raised a further $300 million led by Radical Ventures, with new investor NVIDIA participating, at a reported $4 billion valuation, and reorganized its work around three product lines: the DOS optimization stack, the Lucy world model for immersive experiences, and the Oasis world model for physical AI [26][27]. 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. By 2026 the company had formalized this structure under three named lines: DOS, an inference and training optimization stack licensed to clouds and AI labs; Lucy, its world model for immersive consumer and commerce experiences; and Oasis, its world model aimed at physical AI and robotics simulation [26][28].
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.
In early September 2025, Decart released Oasis 2.0, a successor to the 2024 Oasis demo that the company described as its most advanced real-time world model to that point. Unlike the original Oasis, which generated frames from scratch in response to keyboard and mouse input, Oasis 2.0 is built on a real-time video-to-video model from the MirageLSD lineage: it takes a running game feed and restyles it live into an arbitrary world or art style specified by a text prompt, at 1080p and 30 frames per second [29]. Decart shipped Oasis 2.0 both as a hosted browser demo and as a Minecraft mod distributed through Modrinth and CurseForge, where players install it with the Fabric loader and issue an in-game command and prompt to transform their world into settings such as a different city, a historical era, or a rendering style like anime or pixel art [30]. The mod went through a series of updates over the following months, including bug fixes, additional default worlds, and extended access for creators and streamers [30].
On May 18, 2026, Decart announced a $300 million funding round led by Radical Ventures at a reported $4 billion valuation, bringing the roughly two-year-old company's total disclosed funding to more than $450 million [26][27][31]. The round added a large group of new investors. Strategic and institutional backers included NVIDIA, Atreides Management, Valor Equity Partners, Adobe Ventures, Toyota Ventures, and eBay Ventures, alongside returning investors Sequoia Capital, Benchmark, and Zeev Ventures [26][27]. Individual participants reported by Decart and the press included OpenAI co-founder Andrej Karpathy, former Walt Disney chief executive Michael Eisner, members of the Yamauchi family that founded Nintendo, and gaming investor Moritz Baier-Lentz [27][32]. Several of the new backers, including NVIDIA and Amazon, were described as both investors and commercial customers [26].
Alongside the raise, Decart reorganized its work into three product lines and introduced new flagship versions. The first is DOS (Decart Optimization Stack), an inference and training stack for large language model, agentic, video, and world model workloads that the company says runs across NVIDIA GPUs, Google TPUs, and Amazon Trainium [26][28]. Decart said the new DOS 2.0 delivers more than 1,600 tokens per second for agentic inference, which it framed as roughly eight times a stated industry average of around 200 tokens per second, and supports full-HD video and world model inference at up to 100 frames per second [26][28]. The second line is Lucy, presented as a world model for immersive experiences across gaming, e-commerce, streaming, and advertising, with the company saying it responds to user input in under 30 milliseconds and is already deployed in production for use cases such as virtual try-on, live streaming, and dynamic in-video advertising [26][28]. The third line is Oasis, repositioned as a world model for physical AI that generates three-dimensional environments for applications such as robotics and logistics simulation [28]. Decart said new versions, Lucy 2.5 and Oasis 3, were due in the following weeks but did not give firm dates [26][28].
The 2026 announcement also detailed a partnership with Amazon Web Services. Working with AWS teams including Annapurna Labs, Decart said it became one of the first companies to deploy real-time AI models of this class and scale on Amazon's custom Trainium chips, with its Lucy 2 model running on Trainium3 [26]. Nafea Bshara, a vice president at Annapurna Labs, said Lucy 2 exceeds 80 percent Model FLOPS Utilization on the hardware [26]. AWS's own materials describe Decart's real-time video models on Trainium as reaching up to four times higher frame throughput and two times better cost efficiency than top GPUs, with per-frame latency reduced from about 40 milliseconds to about 10 milliseconds [33]. Decart and Amazon framed the relationship as a commercial and joint go-to-market agreement to bring real-time world models to AWS customers in media, commerce, advertising, and physical AI [26][27].
| 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 |
| 2025 | Oasis 2.0 | Real-time video-to-video world model | Restyles a live game feed from a prompt at 1080p, 30 fps; shipped as a hosted demo and a Minecraft mod |
| 2026 | DOS 2.0 (Decart Optimization Stack) | Inference and training optimization stack | Runs across NVIDIA GPU, Google TPU, and Amazon Trainium; >1,600 tokens/sec agentic inference; up to 100 fps full-HD video and world model inference |
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.
The September 2025 Oasis 2.0 release reworked this approach. Rather than synthesizing the whole game from input alone, Oasis 2.0 runs a live video-to-video model from the MirageLSD lineage over an existing Minecraft feed, restyling it into a prompted world at 1080p and 30 frames per second [29]. Decart distributed it as both a hosted browser demo and a Minecraft mod on Modrinth and CurseForge, letting players transform their world into different places, eras, or art styles in real time [30]. In Decart's 2026 reorganization, the Oasis name was carried forward as the brand for its world model line aimed at physical AI and three-dimensional environment generation, with an Oasis 3 release described as forthcoming [28].
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 first widely covered generation of the line, 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.
By 2026 Decart described Lucy as a world model for immersive experiences and reported that the line was running in production at scale, responding to user input in under 30 milliseconds and serving applications in gaming, e-commerce, streaming, and advertising [26][28]. The company said its Lucy 2 model had been deployed on Amazon's Trainium3 chips and that an updated Lucy 2.5 was due shortly after the May 2026 funding announcement [26][28].
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.
In 2026 Decart formalized this stack under the name DOS (Decart Optimization Stack) and described it as an end-to-end platform spanning hardware-aware model design, kernel tooling, proprietary compilers, and inference optimization [26][28]. The company says the same workload can be retargeted across NVIDIA GPUs, Google TPUs, and Amazon Trainium, compressing what it characterizes as a months-long per-chip optimization process into a few weeks, and reported that DOS 2.0 reaches more than 1,600 tokens per second for agentic inference and up to 100 frames per second of full-HD video and world model inference [26][28]. Decart said the DOS licensing business generates revenue through contracts with several of the world's largest cloud providers, AI labs, and hyperscale companies, without disclosing specific figures [27][28].
As of May 2026, Decart had raised more than $450 million across four disclosed rounds [26][27]. 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 |
| Series C | May 18, 2026 | $300 million | Radical Ventures | NVIDIA, Atreides Management, Valor Equity Partners, Adobe Ventures, Toyota Ventures, eBay Ventures, Sequoia Capital, Benchmark, Zeev Ventures, plus angels (Andrej Karpathy, Michael Eisner, Yamauchi family, Moritz Baier-Lentz) | ~$4 billion |
| Cumulative | May 2026 | >$450 million | - | - | - |
Decart's most prominent investors are Sequoia Capital, Benchmark, Zeev Ventures, Aleph, Radical Ventures, and NVIDIA. The company's Series B was reported by Fortune, SiliconANGLE, Globes, and The SaaS News among others, and its May 2026 Series C was reported by SiliconANGLE, Globes, Calcalist, The Next Web, and Finsmes, among others [26][27][31][32].
| 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 |
| Amazon Web Services | Commercial and go-to-market partnership; Trainium deployment | Working with AWS teams including Annapurna Labs, Decart deployed its Lucy 2 model on Amazon Trainium3 chips and agreed a joint go-to-market to bring real-time world models to AWS customers [26] |
| NVIDIA | Strategic investor and commercial partner | Joined Decart's May 2026 round as a new investor and is described as both an investor and a business partner; DOS targets NVIDIA GPUs among other accelerators [26][27] |
| Sequoia Capital, Benchmark, Zeev Ventures, Aleph, Radical Ventures | Equity investors | Lead and follow-on investors across seed, Series A, Series B, and Series C 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. By 2026, with the DOS, Lucy, and Oasis lines, reporting increasingly described Decart as straddling two markets at once: a real-time creative and commerce video business and a chip-agnostic inference optimization business pitched at clouds, AI labs, and hyperscalers [26][28].
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. The May 2026 Series C drew further coverage from SiliconANGLE, Globes, Calcalist, The Next Web, Verdict, and Finsmes, much of which emphasized NVIDIA's and Amazon's dual role as investors and customers and the company's pitch to make switching between AI chips faster and cheaper [26][27][28][31][32].
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. By 2026 the company described this stack, branded DOS, as hardware-agnostic across NVIDIA GPUs, Google TPUs, and Amazon Trainium, positioning portability across accelerators as a core part of its commercial offering [26][28].
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.