Decart

Decart (also known as Decart AI, operating from the domain decart.ai) is an artificial intelligence company founded in 2023 by three Israeli engineers that builds real-time generative AI systems for video and interactive worlds, along with a chip-agnostic optimization stack for training and running AI models. Its co-founders are chief executive Dean Leitersdorf, Moshe Shalev, and Dean's younger brother Orian Leitersdorf ^[34][38][43]. The company's main engineering office is in Tel Aviv and it also has a San Francisco office; press reports describe it both as Tel Aviv-based and as San Francisco-based with a development center in Tel Aviv ^[4][37]. 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, and in June 2026 it released Oasis 3, an interactive world model for training autonomous vehicles and other robots ^[48].

In August 2025 Decart raised a $100 million Series B at a $3.1 billion valuation, less than a year after emerging from stealth ^[3][5]. In May 2026 it raised a further $300 million led by Radical Ventures, with new investor NVIDIA participating, at a valuation The Wall Street Journal put at nearly $4 billion, 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][38]. Between August and October 2026 the company was the subject of reported acquisition talks: Anthropic discussed buying it for about $6 billion before walking away after due diligence in September, and in October The Wall Street Journal reported that Amazon was in talks to buy it for about $7 billion ^[34][38][44]. 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.

Overview

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 ^[3][4]. Leitersdorf told Calcalist in August 2025 that GPU acceleration was generating "tens of millions of dollars in revenue" ^[4]; the company has not published audited revenue figures, and The Next Web noted in October 2026 that it had not disclosed its revenue ^[34].

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 (first branded Mirage, later the Decart API Platform) 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]. It also operates Cogito, an inference API that serves open-weight large language models ^[65].

History

Founding and Unit 8200 roots (2023)

Decart was registered as a company in Israel on September 7, 2023, about a month before the outbreak of the Israel-Hamas war; both founders were then called up for reserve duty for several months ^[6][42]. The founders are Dean Leitersdorf, who serves as chief executive officer, Moshe Shalev, the chief product officer, and Dean's younger brother Orian Leitersdorf, the company's chief scientist ^[1][34][43]. Announcements and profiles from 2024 and 2025 named only Dean Leitersdorf and Shalev as founders; 2026 coverage and Decart's own media kit list Orian Leitersdorf as the third founder, and Calcalist described him as having largely stayed out of the spotlight ^[1][43][72]. Dean Leitersdorf and 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; Calcalist reported that their partnership began with a hallway conversation at their base in 2021 ^[6]. Dean Leitersdorf, who was in his mid-twenties at founding, holds three computer science degrees including a PhD completed at age 23 at the Technion ^[6]. Zeev Ventures became the company's first investor in September 2023 ^[1]. Profiles in Calcalist describe a demanding work culture centered on the company's Tel Aviv office ^[6].

According to Calcalist, the company's first product was software that increased the efficiency of NVIDIA GPUs, which won a multimillion-dollar customer by December 2023 and gave Decart revenue before it had a consumer product ^[6].

Stealth exit and seed round (October 2024)

Decart emerged from stealth on October 31, 2024, announcing a $21 million seed round led by Sequoia Capital with participation from Zeev Ventures ^[1]. Calcalist reported that the seed round itself had been raised in August 2024 ^[6], and TechCrunch, citing sources, reported that it valued the company at just over $100 million ^[2]. The announcement described Decart as an efficiency-focused AI research lab whose infrastructure platform improved the efficiency of training and inference for large foundation models, and said it had reached early profitability within a few months of launch by selling that platform to paying enterprise customers ^[1]. Sequoia partner Shaun Maguire compared the team's low-level systems work to early Google pushing commodity hardware to its limits ^[1].

Oasis and Etched partnership (October to November 2024)

On October 31, 2024, Decart and Etched released Oasis, described by the two companies as the first playable, real-time, open-world AI model ^[10][56]. 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 ^[9][10].

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 ^[56]. 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 ^[10]. 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 Decart said the hosted demo passed one million users within three days ^[6][12][13]. Calcalist reported that Elon Musk posted "wow" on X after seeing it ^[6].

Benchmark round and pivot to real-time video (December 2024)

On December 19, 2024, TechCrunch reported that Decart had raised a $32 million Series A led by Benchmark, less than two months after announcing its seed, and that, according to its sources, the post-money valuation was more than $500 million ^[2]. The article framed the round as funding for "AI tech and open world apps," and described Decart as based in San Francisco with operations in Israel. Leitersdorf told TechCrunch that the company had already been approached by would-be acquirers several times and that its goal was a trillion-dollar company ^[2]. After this round Decart had raised approximately $53 million in total ^[6][22].

MirageLSD and Crusoe Cloud distribution (mid 2025)

On July 17, 2025, Decart introduced MirageLSD, which it described as the first Live Stream Diffusion model ^[14]. 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 with a response time under 40 milliseconds ^[14]. 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 ^[14]. 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 ^[15].

Technion partnership (July 2025)

On July 7, 2025, Decart and the Technion announced that Decart would fund the Technion's excellence program, renamed the Technion-Decart Excellence Program, with millions of shekels over the following years. Dean Leitersdorf and seven other Decart employees are graduates of the program. The two institutions also signed a memorandum of understanding to set up a joint AI research center ^[22].

Series B at a $3.1 billion valuation (August 2025)

On August 7, 2025, Decart announced a $100 million Series B with participation from existing investors Sequoia Capital, Benchmark, and Zeev Ventures and new investor Aleph; the reports did not name a lead investor ^[3][8]. Fortune reported that the round valued the company at $3.1 billion, though Decart did not comment on specific numbers, and put Decart's cumulative raise at about $153 million ^[3]; Globes described the valuation as a more than six-fold increase on the December 2024 round ^[7]. SiliconANGLE and Globes both reported the round as focused on scaling Decart's real-time video platform ^[7][8]. Leitersdorf said the company had spent less than $10 million of its investors' money because revenue covered its GPU costs, and Calcalist reported that headcount had grown over the previous year from 15 to more than 60, mostly in Tel Aviv, with additional teams in northern Israel, San Francisco, and New York ^[3][4]. With the round, Decart opened a San Francisco research and development center led by Kfir Aberman, formerly of Snap and Google and a co-creator of DreamBooth ^[5].

Oasis 2.0 (September 2025)

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].

Lucy Edit and the developer platform (September to November 2025)

On September 18, 2025, Decart released the weights and reference code of Lucy Edit Dev, an open-weight model of about 5 billion parameters for instruction-guided video editing, built on the Wan2.2 5B architecture and published on Hugging Face under a non-commercial license ^[53]. It edits existing clips from plain text instructions, such as changing clothing, replacing a character, or swapping the scene, while preserving the original motion; Decart described it as the first open-source instruction-guided video editing model ^[53]. A Lucy Edit 1.1 Dev checkpoint followed in November 2025 ^[71]. In October 2025 Decart added real-time lip syncing to its API and published a reference application that paired its lip-sync service with ElevenLabs voices to animate talking characters ^[51][67]. In November 2025 the Decart API Platform became available on the AWS Marketplace with integrated billing ^[51].

Trainium3 at AWS re:Invent (December 2025)

In 2025 Amazon Web Services gave Decart early access to its Trainium3 chip after meeting the founders, The Wall Street Journal later reported ^[34][35]. When Amazon announced its Trainium3 UltraServers on December 2, 2025, at re:Invent, it named Decart as a customer achieving "4x faster frame generation at half the cost of GPUs" on the chip ^[55]. During the conference, Decart's real-time video model ran live on Trainium3 on the keynote stage, turning AWS senior vice president Peter DeSantis into a cartoon of Amazon CTO Werner Vogels, according to GeekWire ^[37].

Lucy 2.0 (January 2026)

Following MirageLSD, Decart introduced the Lucy line of real-time video transformation models. Lucy 2.0, released on January 26, 2026, 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 ^[18][19][20]. 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 ^[18]. The Decoder reported that it ran on AWS Trainium3 chips ^[19]. Decart said sustained real-time generation with Lucy 2.0 cost roughly three dollars per hour of generated video ^[20].

AI Grid tests with Comcast and NVIDIA (March 2026)

On March 17, 2026, Decart said it was working with Comcast and NVIDIA on an "AI Grid," a distributed infrastructure that moves inference from centralized data centers to the network edge. Decart reported that running Lucy on NVIDIA GPUs at an edge node inside Comcast's network cut round-trip network time to about 20 milliseconds and held a steady 29 to 30 frames per second ^[54]. The same month, Decart renamed its Mirage model to Lucy Restyle Live across its API and SDKs, and in April 2026 it released Lucy 2.1 together with a real-time virtual try-on model ^[51].

$300 million round, the DOS platform, and AWS partnership (May 2026)

On May 18, 2026, Decart announced a $300 million funding round led by Radical Ventures, bringing the roughly two-year-old company's total disclosed funding to more than $450 million ^[26][27][31]. Decart did not disclose the valuation; The Wall Street Journal reported that the round valued the company at nearly $4 billion, up from $3.1 billion in August 2025 ^[28][32][38]. Decart did not give the round a series letter; Canadian outlet The Logic called it a Series B and, citing PitchBook data, the second-largest round Radical had led or co-led ^[57]. 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][31][32]. Decart described NVIDIA as both an investor and a business partner and Amazon as a strategic customer; Amazon was not listed among the investors ^[26][27].

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]. A statement by Leitersdorf on AWS's Trainium customer page says early testing of Decart's real-time video models on Trainium showed up to four times higher frame throughput and two times better cost efficiency than top GPUs, with 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].

Oasis 3 and Lucy 2.5 (June to July 2026)

On June 10, 2026, Decart released Oasis 3, which it called the first interactive world model for physical AI and the first world model available through an API from launch ^[48]. The model generates driving environments from a text prompt, is conditioned on the actions of a robot or vehicle rather than keyboard and mouse input, and outputs three synchronized camera views (one front-facing and two side-facing) ^[48][49]. Decart said it runs at 22 frames per second, at a resolution it gives as 512x768x3, with under 200 milliseconds of latency, on CoreWeave's cloud and NVIDIA HGX B200 systems ^[48]. The first target market is autonomous vehicle developers who need to train and test against rare scenarios; Decart offered it at $0.02 per second of generation, with a Python SDK and a notebook for reinforcement learning inside the simulator ^[49][51][52]. TechCrunch reported that Decart had a community of more than 100,000 developers, many building on Lucy ^[49].

On July 16, 2026, Decart released Lucy 2.5, which it said improved edit quality, physically plausible visual effects such as water, fire, and sand, object removal, and temporal consistency while still generating 1080p video at 30 frames per second ^[50]. In October 2026 it added Lucy 2.6 to its API, along with updated virtual try-on models ^[51].

Acquisition talks (August to October 2026)

In August 2026 Israeli business media reported that Decart was negotiating a sale. On August 10, Globes reported offers estimated at about $5 billion, with NVIDIA believed to be among the bidders, while Calcalist reported advanced talks at about $6 billion to $7 billion and said that NVIDIA's earlier advanced negotiations had been set aside after a larger bid emerged ^[39][40]. Calcalist said high-tech sources speculated the new bidder could be SpaceX and that Amazon and Nebius were also interested; Elon Musk called the SpaceX report "fake news" and said SpaceX would not acquire Decart ^[38][39][41].

On August 13, 2026, Bloomberg reported that Anthropic was in talks to buy Decart for about $6 billion, which would have been Anthropic's largest known acquisition. According to Bloomberg's sources, Decart's team would join Anthropic's inference and performance organization, and its chip-efficiency software could help Anthropic's existing infrastructure absorb more demand for Claude ^[38][41]. Calcalist reported on August 16 that the parties were exchanging advanced drafts of an agreement valuing Decart at about $7 billion, mostly payable in Anthropic shares, and that the founders and Sequoia, Decart's largest investor, preferred Anthropic even though NVIDIA was believed to have offered a higher valuation ^[42]. Globes reported that the founders had rejected the higher NVIDIA offer in favor of Anthropic's share offer in anticipation of Anthropic's IPO ^[44].

On September 8, 2026, Bloomberg reported that Anthropic had decided against the acquisition after conducting due diligence, adding that the companies might still collaborate in other ways; both declined to comment ^[44][45][47]. Neither company gave a reason. Globes, citing sources close to Decart, reported that the founders had refused Anthropic's push to move much of the team from Tel Aviv to San Francisco ^[34][37]. In an analysis, Calcalist's Sophie Shulman raised other possible explanations, including whether the efficiency gains held at very large scale and whether Anthropic doubted the founders' long-term commitment after a sale ^[46].

On October 11, 2026, The Wall Street Journal reported that Amazon was in talks to acquire Decart in a deal that could value it at about $7 billion, and that a deal could come together soon unless talks fell apart or another buyer emerged ^[34][35]. Amazon declined to comment outside business hours and Decart did not respond to requests for comment, Reuters reported ^[34]. Globes calculated that the reported price was about 75 percent above the May 2026 valuation and said a deal would be among the largest exits in Israeli technology history ^[36]. As of the mid-October reports, neither company had confirmed an agreement ^[35][37].

Products and technology

Product timeline

Oasis

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. The model was trained with Diffusion Forcing, which denoises with independent per-token noise levels ^[10].

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 ^[10][56]. Decart released the weights for a 500 million parameter Oasis model and the inference code, while keeping larger checkpoints behind the hosted demo ^[10]. 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 ^[11][13].

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 ^[28].

Oasis 3 (June 2026) generates worlds autoregressively, frame by frame, with each frame conditioned on the previous frames and the vehicle's latest action, which closes the loop needed to train a policy inside the model. Decart says a geometry-aware architecture keeps its three camera views consistent with each other, and that it was trained on real-world data, including the NVIDIA Physical AI Open Dataset on Hugging Face, and post-trained for realism ^[48]. Decart lists online reinforcement learning, policy evaluation, and teleoperation for collecting expert demonstrations as uses, and says it plans to extend the model from autonomous vehicles to humanoids, manipulation robots, and off-road systems ^[48]. In a hands-on review, TechCrunch's Rebecca Bellan found Oasis 3 produced the most photorealistic environments from a single prompt of the world models she had tried, including Genie 3 and Marble, but that scenes drifted away from the prompt as she drove, a starting intersection was gone when she turned back, the controls were not very responsive, and the car could drive through other cars. Leitersdorf called the physics problem "a major research problem that we're cracking now" ^[49].

MirageLSD (Live Stream Diffusion)

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 a response time under 40 milliseconds, 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 ^[14]. Decart positions MirageLSD as an enabling layer for live streaming, virtual production, and interactive media; VP Land reported that it also supported "vibe-coding" of simple games in which the model generates the visual style ^[17]. In March 2026 Decart renamed the Mirage model Lucy Restyle Live in its API, folding it into the Lucy family ^[51].

Lucy

The Lucy family extends MirageLSD's real-time video work into a more general purpose model suitable for higher resolution and longer sessions. Decart's first Lucy release in 2025 introduced localized, frame-by-frame video editing ^[24]; a Decart engineer later told AI World Today that the original model struggled to keep characters consistent in clips longer than about 30 seconds ^[21]. Lucy 2.0, which Decart said builds on MirageLSD and Lucy Edit 1.0, runs at 30 frames per second at 1080p resolution, with sub-100 millisecond per-frame latency ^[18]. 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 ^{[18][19][20][21]}. 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 ^[18]. To limit drift, Decart trains the model on its own imperfect outputs with a method it calls Smart History Augmentation, and it pitched Lucy 2.0 for robotics data augmentation as well as media ^[18]. Decart has said sustained generation costs roughly $3 per hour of output ^[20].

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 ^[26]. Lucy 2.5 (July 2026) added a technique Decart calls Self-Anchoring, in which the model adopts a snapshot of its own edited output early in a stream as its reference so that edits such as a costume change stay fixed for minutes, and was trained on a synthetic video editing corpus made by propagating image edits across video sequences ^[50]. To keep latency down, Decart said it runs Lucy 2.5 in MXFP8 and NVFP4 precision, uses a custom dynamic sparse attention algorithm, and fuses kernels more aggressively ^[50]. Lucy 2.6 followed in the API in October 2026; real-time Lucy sessions are billed at $0.02 per second, or $0.04 per second in a lower-latency fast mode ^[51][52].

Developer platform, DOS and Cogito

Decart first exposed its models through an API platform branded Mirage at mirage.decart.ai. By 2026 its developer surface was the Decart API Platform at platform.decart.ai, with JavaScript, Python, Swift, and Android SDKs, real-time and batch endpoints for Lucy editing, restyling, and virtual try-on, and per-second pricing ^[25][51][52]. Decart has also partnered with Crusoe Cloud to make MirageLSD available within Crusoe's GPU cloud environment for enterprise customers ^[15]. 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 ^[3][4]. Decart has cited cost reductions for video generation from tens to thousands of dollars per hour of output down to under $0.25 per hour, though specifics of those benchmarks have not been independently verified ^[4][7].

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]. Its product page says it works with hyperscalers, chip manufacturers, and AI labs on workloads across GPUs, TPUs, Trainium, AMD chips, and other accelerators ^[68]. 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]. Calcalist reported in August 2026 that Decart's revenue was estimated at several tens of millions of dollars, with uneven growth driven largely by projects for chip manufacturers ^[39].

Decart also sells LLM inference directly through Cogito, an API that serves open-weight models including gpt-oss, Kimi K2.6, Kimi K2.7 Code, Qwen, and GLM-5.2, and which it markets to builders of latency-sensitive AI agents ^[65]. In June 2026 Decart published an NVFP4-quantized version of Kimi K2.7 Code on Hugging Face, stating that the model was available through Cogito and through OpenRouter ^[66].

Research

Decart publishes some of its research. A paper on Adaptive-Origin Guidance (AdaOr), a method that turns the strength of a text-driven image or video edit into a continuous slider, was posted to arXiv in February 2026 by authors from Decart and Tel Aviv University, including Kfir Aberman, and Decart lists it as a SIGGRAPH 2026 paper ^[64][69]. In October 2026 Decart published a technical report on a learned spatial memory codec that lets a video world model store and read back compact memories of journeys lasting nearly two hours ^[69]. The Lucy Edit release was accompanied by a technical paper on text-guided video editing ^[53].

Funding

As of October 2026, Decart had raised more than $450 million across four disclosed rounds ^[26][27][42]. The company is privately held and its detailed cap table has not been publicly released.

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 round was reported by SiliconANGLE, Globes, Calcalist, The Next Web, and Finsmes, among others ^{[26][27][31][32]}. In June 2026 Leitersdorf told TechCrunch that the company had burned "drastically less" than $100 million over its lifetime ^[49].

Reported acquisition talks

None of the following talks had produced an announced deal as of the latest reports in mid-October 2026.

Leadership and people

Sources for the table: ^{[2][5][6][43][58][60]}. 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. Calcalist reported in August 2026 that the company employed about 100 people ^[41][42].

Founders' research background

Dean Leitersdorf's academic work was in distributed computing. His Technion doctoral thesis, "Fast Distributed Algorithms via Sparsity Awareness," supervised by Keren Censor-Hillel, was one of two recipients of the 2023 Principles of Distributed Computing Doctoral Dissertation Award ^[58][59]. His papers with Censor-Hillel and others address problems such as sparse matrix multiplication and triangle listing in the congested clique model of networked computation, and argue that algorithms which are aware of the sparsity of their input can run faster across several distributed settings ^[61][70]. Calcalist reported that his doctoral work led to a postdoc at the National University of Singapore, where he was working when ChatGPT was released ^[6].

Orian Leitersdorf's doctoral research at the Technion's Viterbi Faculty of Electrical and Computer Engineering focused on components for next-generation computers ^[60]. His papers, mostly with Shahar Kvatinsky and Ronny Ronen, concern digital processing-in-memory, in which arithmetic is performed inside memristive memory arrays to avoid moving data to a processor; they include PyPIM, which connects such hardware to Python tensor code ^[62]. The brothers co-authored AritPIM, a 2022 paper on high-throughput arithmetic for processing-in-memory ^[63].

Partnerships

Decart has signed a number of partnerships that combine its models with third-party hardware, cloud, or research infrastructure.

Position in the AI video market

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.

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 ^[16][19][28]. 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]. With Oasis 3, TechCrunch placed it in an "increasingly packed world model arena" alongside Google DeepMind's Genie 3, World Labs' Marble, and world-model efforts at Luma and Runway ^[49]. The acquisition interest of 2026 centered mostly on the second business: Bloomberg and The Next Web described the chip-efficiency software as the main attraction for Anthropic, while The Wall Street Journal and GeekWire pointed to the chip-switching software as the reason for Amazon's interest, with GeekWire noting that the video models could also fit Amazon businesses such as Twitch and online retail ^{[34][37][38][45]}.

Industry coverage and reception

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." Calcalist ranked Decart first on its list of the 50 most promising Israeli startups for 2025 ^[6]. Ynet has described Decart as an Israeli AI unicorn aiming at large entertainment markets, comparing its ambitions to Netflix, YouTube, and TikTok ^[23]. The May 2026 funding round drew further coverage from SiliconANGLE, Globes, Calcalist, The Next Web, Verdict, and Finsmes, much of which emphasized NVIDIA's dual role as investor and partner, Amazon's role as a strategic customer, 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. TechCrunch's June 2026 review of Oasis 3 found the same pattern at higher fidelity: photorealistic opening scenes that lost coherence over a long drive ^[49].

Business coverage in 2026 was more skeptical than the earlier funding stories. Globes cited market sources who doubted acquisition interest because of Decart's low sales and its changes of product focus, and who thought buyers were more interested in its concentration of engineers ^[40]. Calcalist wrote that after three years Decart still lacked a single clearly defined mass-market product or a recurring revenue engine, earning money largely through projects, and that the scale of its business remained far smaller than the valuations being discussed for a sale ^[35][42].

Technology and approach

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 ^[14]. Later models extended the idea: Lucy 2.0's Smart History Augmentation trains on the model's own imperfect outputs, and Lucy 2.5's Self-Anchoring uses an early frame of the model's own output as a fixed reference ^[18][50]. 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 and the Lucy Edit Dev model while keeping flagship models hosted ^[10][53]. 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. Its main engineering office is in Tel Aviv, and it maintains close ties to Israeli universities, especially the Technion ^[22].

Continue reading

• WWorld modelA world model is an artificial intelligence system that learns an internal representation of how an environment works, enabling it to predict future states, simulate the consequences of actions, and support…Mentioned 28 times above
• NvidiaNvidia Corporation is an American technology company that designs graphics processing units (GPUs) and is the world's dominant supplier of hardware for artificial intelligenceMentioned 28 times above
• EtchedEtched is an American semiconductor startup building Sohu, the first application-specific integrated circuit (ASIC) designed to run only transformer models, the architecture behind most large language models.Mentioned 11 times above
• Runway (company)Runway is a New York based artificial intelligence company that builds generative AI tools for video, with its Gen series of models and its GWM-1 world models.Mentioned 4 times above
• Luma Dream MachineLuma Dream Machine is a generative AI video and image platform from Luma AI (Luma Labs, Inc.) that turns text prompts and still images into short, realistic video clips.Mentioned 3 times above
• Pika 2.5Pika 2.5 is a generative video model developed by Pika Labs, the San Francisco based AI video startup co-founded in April 2023 by Stanford AI Lab dropouts Demi Guo and Chenlin Meng.Mentioned 3 times above

More from See also

• List of neolabs
• Neolab
• Hedra Character
• Generative AI
• Diffusion model
• Genie 3
• World Labs
• Odyssey (AI lab)
• NVIDIA Cosmos
• Amazon Web Services
• AWS Trainium 3

Topics:AI CompaniesGenerative AIVideo Generation

References

1. ^1 ^2 ^3 ^4 ^5 ^6"Decart Emerges From Stealth With $21 Million Seed Funding to Enhance AI Efficiency, Bringing Costly AI Within Reach." Access Newswire, October 31, 2024. accessnewswire.com/...seed-funding-to-enhan-937450
2. ^1 ^2 ^3 ^4"Decart nabs $32M at $500M+ valuation to build AI tech and 'open world' apps." TechCrunch, December 19, 2024. techcrunch.com/...-another-32m-at-a-500m-valuation
3. ^1 ^2 ^3 ^4 ^5 ^6"Exclusive: Decart raises $100 million at a $3.1 billion valuation, chasing the future of real-time creative AI." Fortune, August 7, 2025. fortune.com/...the-future-of-real-time-creative-ai
4. ^1 ^2 ^3 ^4 ^5 ^6"Decart hits $3.1 billion valuation on $100 million raise to power real-time interactive AI." CTech (Calcalist), August 7, 2025. calcalistech.com/...m7yk4c4e3
5. ^1 ^2 ^3"From stealth to $3.1 billion in less than a year." CTech (Calcalist), August 7, 2025. calcalistech.com/...kqp6pwozq
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Sophie Shulman. "Inside Decart: The Israeli startup chasing AI greatness, one sleepless night at a time." CTech (Calcalist), April 30, 2025. calcalistech.com/...6rn7vscw2
7. ^1 ^2 ^3"Israeli AI co Decart raises $100m at $3.1b valuation." Globes, August 7, 2025. en.globes.co.il/...00m-at-31b-valuation-1001518392
8. ^1 ^2"Decart raises $100M on $3.1B valuation to grow real-time AI video platform." SiliconANGLE, August 7, 2025. siliconangle.com/...ow-real-time-ai-video-platform
9. ^"Decart and Etched Release Oasis, a New AI Model Transforming Gaming Worlds." InfoQ, November 2024. infoq.com/...decart-etched-oasis
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7"Oasis: A Universe in a Transformer." Oasis project page, Decart and Etched, 2024. oasis-model.github.io
11. ^"Oasis (Minecraft clone)." Wikipedia. en.wikipedia.org/...Oasis_(Minecraft_clone)
12. ^"I've seen the future of gaming, AI model creates playable Minecraft from a picture." Tom's Guide, 2024. tomsguide.com/...playable-minecraft-from-a-picture
13. ^1 ^2"Oasis is an interactive AI experience that turns Minecraft into a nightmarish hallucination." TechSpot, 2024. techspot.com/...inecraft-nightmarish-hallucination
14. ^1 ^2 ^3 ^4 ^5"MirageLSD: The First Live-Stream Diffusion AI Video Model." Decart AI publications, July 17, 2025 (archived copy; the original page has been removed). web.archive.org/...mirage
15. ^1 ^2 ^3"MirageLSD: Decart's real-time AI video model is now available on Crusoe Cloud." Crusoe AI blog. crusoe.ai/...odel-is-now-available-on-crusoe-cloud
16. ^"Mirage AI Instantly Transforms Live Video With Just a Text Prompt." eWeek. eweek.com/...mirage-ai-live-video-rea-time-changes
17. ^"Real-Time AI Video Arrives with MirageLSD's Live-Stream Diffusion Technology." VP Land. vp-land.com/...-s-live-stream-diffusion-technology
18. ^1 ^2 ^3 ^4 ^5 ^6 ^7"Introducing Lucy 2: SOTA realtime world transformation model." Decart AI publications, January 26, 2026. decart.ai/...ing-sota-video-generation-in-realtime
19. ^1 ^2 ^3 ^4"Decart's Lucy 2.0 transforms live video in real time using text prompts." The Decoder, January 28, 2026. the-decoder.com/...in-real-time-using-text-prompts
20. ^1 ^2 ^3 ^4"Live AI video marketing with Decart's new Lucy 2 model." Marketing Tech News, February 2, 2026. marketingtechnews.net/...-decarts-new-lucy-2-model
21. ^1 ^2"With Decart's New Model, Real-time Video Transformation Just Got Real." AI World Today, February 2, 2026. aiworldtoday.net/...with-decarts-new-model-real-time
22. ^1 ^2 ^3 ^4Meir Orbach. "Decart pledges millions to Technion in strategic AI push." CTech (Calcalist), July 7, 2025. calcalistech.com/...sjca2dfrll
23. ^"The Israeli AI unicorn that wants to replace Netflix, YouTube and TikTok." Ynetnews. ynetnews.com/...hy4mrdfdll
24. ^"The Live AI Lab" (About page, including company timeline). Decart. decart.ai/about
25. ^Decart API Platform. platform.decart.ai
26. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24"Decart Raises $300M: Tech Leaders Back the Company as Both Customers and Investors." Decart AI publications, May 18, 2026. decart.ai/...mpany-as-both-customers-and-investors
27. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11"Nvidia backs Israeli AI unicorn Decart in $300 million funding round at $4 billion valuation." CTech (Calcalist), May 18, 2026. calcalistech.com/...sjt9ncukgl
28. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16"Decart raises $300M for its AI optimization software, world models." SiliconANGLE, May 18, 2026. siliconangle.com/...mization-software-world-models
29. ^1 ^2"Introducing Oasis 2.0: Our most advanced AI model. Transform game worlds and styles in real time. 1080p, 30fps." Decart AI (DecartAI) post on X, September 4, 2025. x.com/...1963758685995368884
30. ^1 ^2 ^3"Oasis 2.0." Modrinth mod page, Decart AI, 2025. modrinth.com/...oasis2
31. ^1 ^2 ^3 ^4"Decart raises $300m including Nvidia investment." Globes, May 18, 2026. en.globes.co.il/...ia-investment-report-1001543260
32. ^1 ^2 ^3 ^4"Decart raises $300M to put a real-time world model in front of Amazon's chips." The Next Web, May 18, 2026. thenextweb.com/...on-radical-ventures-world-models
33. ^"AI Accelerator - AWS Trainium Customers." Amazon Web Services. aws.amazon.com/...customers
34. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Cristian Dina. "Amazon in talks to buy Decart for about $7bn, WSJ reports." The Next Web, October 13, 2026. thenextweb.com/...decart-7bn-acquisition-talks-wsj
35. ^1 ^2 ^3 ^4 ^5Sophie Shulman and Golan Hazani. "Amazon emerges as potential buyer of Israeli AI startup Decart at $7 billion." CTech (Calcalist), October 12, 2026. calcalistech.com/...ec25jzet0
36. ^1 ^2Meytal Vaizberg. "Amazon in talks to buy Israel's Decart at $7b valuation - report." Globes, October 12, 2026. en.globes.co.il/...cart-at-7b-valuation-1001559158
37. ^1 ^2 ^3 ^4 ^5 ^6Todd Bishop. "Why Amazon might pay billions for Decart, a 3-year-old AI startup that hasn't disclosed its revenue." GeekWire, October 12, 2026. geekwire.com/...p-that-hasnt-disclosed-its-revenue
38. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Shirin Ghaffary (Bloomberg). "Anthropic said in talks to buy startup Decart for $6 billion." Fortune, August 13, 2026. fortune.com/...to-buy-startup-decart-for-6-billion
39. ^1 ^2 ^3 ^4 ^5Golan Hazani and Sophie Shulman. "Israeli AI startup Decart nearing $6 billion sale, Musk's SpaceX circling." CTech (Calcalist), August 10, 2026. calcalistech.com/...hjhzrluuml
40. ^1 ^2Assaf Gilead. "Israeli AI co Decart mulls $5b exit." Globes, August 10, 2026. en.globes.co.il/...decart-mulls-5b-exit-1001551840
41. ^1 ^2 ^3 ^4Golan Hazani and Sophie Shulman. "Anthropic in talks to acquire Israeli AI startup Decart for $6 billion." CTech (Calcalist), August 13, 2026. calcalistech.com/...mrrffazk1
42. ^1 ^2 ^3 ^4 ^5 ^6 ^7Golan Hazani and Sophie Shulman. "Anthropic closes in on $7 billion Decart deal after beating Nvidia to the table." CTech (Calcalist), August 16, 2026. calcalistech.com/...b1evv3aufg
43. ^1 ^2 ^3 ^4Allon Sinai. "Anthropic's $7 billion Decart deal puts a spotlight on its little-known 23-year-old co-founder and chief scientist." CTech (Calcalist), August 16, 2026. calcalistech.com/...hkbg4d1wgl
44. ^1 ^2 ^3 ^4 ^5Assaf Gilead. "Anthropic decides against $6b Decart acquisition - report." Globes, September 8, 2026. en.globes.co.il/...t-acquisition-report-1001554747
45. ^1 ^2 ^3Ana-Maria Stanciuc. "Anthropic has walked away from its $6bn Decart deal, Bloomberg reports." The Next Web, September 8, 2026. thenextweb.com/...alks-away-decart-6bn-acquisition
46. ^Sophie Shulman. "What made Anthropic walk away from the $6 billion Decart deal?" CTech (Calcalist), September 8, 2026. calcalistech.com/...sy77he6oml
47. ^"Anthropic drops plans to acquire Israeli AI startup Decart." CTech (Calcalist), September 8, 2026. calcalistech.com/...qs4rvoqbv
48. ^1 ^2 ^3 ^4 ^5 ^6 ^7"Introducing Oasis 3: First Interactive World Model for Physical AI." Decart AI publications, June 10, 2026. decart.ai/...teractive-world-model-for-physical-ai
49. ^1 ^2 ^3 ^4 ^5 ^6 ^7Rebecca Bellan. "Decart's new world model can simulate hours of photorealistic driving, with some caveats." TechCrunch, June 10, 2026. techcrunch.com/...listic-driving-with-some-caveats
50. ^1 ^2 ^3 ^4"Lucy 2.5: Raising the Bar for Live AI." Decart AI publications, July 16, 2026. decart.ai/...lucy-2-5-raising-the-bar-for-live-ai
51. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9"Changelog." Decart API Platform documentation, accessed October 27, 2026. docs.platform.decart.ai/changelog
52. ^1 ^2 ^3"Pricing." Decart API Platform documentation, accessed October 27, 2026. docs.platform.decart.ai/...pricing
53. ^1 ^2 ^3 ^4"Lucy Edit Dev (5B)." Model card, decart-ai on Hugging Face. huggingface.co/...Lucy-Edit-Dev
54. ^1 ^2"Closing the Latency Gap: Decart Is Helping Shape the AI Grid." Decart AI publications, March 17, 2026. decart.ai/...decart-is-helping-shape-the-ai-grid
55. ^"Trainium3 UltraServer delivers faster AI training at lower cost." About Amazon, December 2, 2025. aboutamazon.com/...r-faster-ai-training-lower-cost
56. ^1 ^2 ^3"Oasis: A Universe in a Transformer." Decart AI publications, October 31, 2024. decart.ai/...oasis-interactive-ai-video-game-model
57. ^"Decart's US$300M fundraise is the latest big AI bet for Toronto's Radical Ventures." The Logic, May 19, 2026. thelogic.co/...i-bet-for-torontos-radical-ventures
58. ^1 ^2"PODC23 Outstanding Doctoral Award to Dean Leitersdorf." Technion Henry and Marilyn Taub Faculty of Computer Science, March 28, 2023. cs.technion.ac.il/...view-new
59. ^"Doctoral Dissertation Award." International Symposium on Distributed Computing (DISC). disc-conference.org/...doctoral-dissertation-award
60. ^1 ^2"You Are a Source of Inspiration for Us." Technion - Israel Institute of Technology, September 15, 2025. technion.ac.il/en
61. ^Keren Censor-Hillel, Dean Leitersdorf, and Volodymyr Polosukhin. "On Sparsity Awareness in Distributed Computations." arXiv:2105.06068, 2021. arxiv.org/...2105.06068
62. ^Orian Leitersdorf, Ronny Ronen, and Shahar Kvatinsky. "PyPIM: Integrating Digital Processing-in-Memory from Microarchitectural Design to Python Tensors." arXiv:2308.14007, 2023. arxiv.org/...2308.14007
63. ^Orian Leitersdorf, Dean Leitersdorf, Jonathan Gal, Mor Dahan, Ronny Ronen, and Shahar Kvatinsky. "AritPIM: High-Throughput In-Memory Arithmetic." arXiv:2206.04218, 2022. arxiv.org/...2206.04218
64. ^Alon Wolf, Chen Katzir, Kfir Aberman, and Or Patashnik. "Continuous Control of Editing Models via Adaptive-Origin Guidance." arXiv:2602.03826, 2026. arxiv.org/...2602.03826
65. ^1 ^2"Cogito: Ultra-fast Inference." Decart, accessed October 27, 2026. cogito.decart.ai
66. ^"Kimi-K2.7-Code-NVFP4." Model card, decart-ai on Hugging Face, June 2026. huggingface.co/...Kimi-K2.7-Code-NVFP4
67. ^1 ^2"Building Living Characters With Decart and ElevenLabs." Decart AI publications, October 22, 2025. decart.ai/...characters-with-decart-and-elevenlabs
68. ^"Decart Optimization Stack." Decart. decart.ai/optimization-engine
69. ^1 ^2"Blog." Decart, accessed October 27, 2026. decart.ai/blog
70. ^Keren Censor-Hillel, Dean Leitersdorf, and Elia Turner. "Sparse Matrix Multiplication and Triangle Listing in the Congested Clique Model." arXiv:1802.04789, 2018. arxiv.org/...1802.04789
71. ^"Lucy Edit 1.1 Dev (5B)." Model card, decart-ai on Hugging Face, November 2025. huggingface.co/...Lucy-Edit-1.1-Dev
72. ^"Newsroom" (including media kit). Decart, accessed October 27, 2026. decart.ai/newsroom

Show all 72 references