Haiper
Last reviewed
Jun 4, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,879 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 4, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,879 words
Add missing citations, update stale details, or suggest a clearer explanation.
Haiper was a London-based generative AI startup that built consumer video-generation models, operating the haiper.ai web app between 2024 and early 2025. Founded in late 2021 by former Google DeepMind researchers Yishu Miao (CEO) and Ziyu Wang (CTO), the company emerged from stealth in March 2024 with a $13.8 million seed round and positioned itself as a European challenger to text-to-video tools such as Runway and OpenAI's Sora. Haiper described its longer-term ambition as a "perceptual foundation model" able to understand the physics of the visual world. The consumer product was short-lived: the web app went dark in early February 2025, both co-founders left for Microsoft AI in March 2025, and by June 2025 the company's video-generation models had been acquired by another firm, NetMind.AI, effectively ending Haiper as an independent product.[1][2][3][4]
Haiper was founded by Dr. Yishu Miao and Dr. Ziyu Wang. Both hold PhDs in machine learning from the University of Oxford and previously worked as researchers at Google DeepMind. Miao, who served as chief executive, had also worked on TikTok's global trust and safety team before founding the company; Wang, the chief technology officer, had been a research scientist at DeepMind and Google.[1][2]
The pair began working on the company in 2021 and formally incorporated it in 2022, with offices in the King's Cross area of London. The startup spent roughly two years in stealth, during which Miao said the team built distributed data-processing and model-training infrastructure intended to scale, and shifted its focus specifically to video generation in the months before its public launch.[1][2][5]
Haiper framed its mission in unusually expansive terms for a video tool. Miao described the goal as building "an AGI with full perceptual abilities, which has boundless potential to assist with creativity," and said the company's visual foundation model would be "a leap forward in AI's ability to deeply understand the physics of the world" and replicate the essence of reality in the videos it generated. The company positioned its products as steps toward this "perceptual foundation model" rather than as ends in themselves.[2][5]
Haiper disclosed two financing rounds during its independent life, totaling $19.2 million.[1][6]
| Round | Date announced | Amount | Lead / investors |
|---|---|---|---|
| Pre-seed | April 2022 | $5.4 million | Angel investors including Phil Blunsom and Nando de Freitas |
| Seed | March 2024 | $13.8 million | Led by Octopus Ventures, with participation from 5Y Capital |
The seed round, reported by TechCrunch and others on March 5, 2024, coincided with the company's emergence from stealth. Coverage put Haiper's total disclosed funding at $19.2 million and noted that the company was expected to pursue a Series A round during 2024. A Series A was never publicly announced.[1][6]
Some UK reporting expressed the figures in pounds sterling, citing a roughly £10.8 million seed round and total funding of about £15.1 million, which reflect currency conversion of the same dollar amounts rather than separate financings. Filings reported by Sifted via Companies House later indicated that Haiper's cash position had fallen to about £3.5 million as of July 31, 2024, down from roughly £10.9 million a year earlier, an indication of the capital-intensive nature of training video models.[3][7]
Among the pre-seed angels, Nando de Freitas is notable in hindsight: he later became a vice president of AI at Microsoft, and the Haiper founders joined his organization there in 2025.[3][4]
Haiper's consumer offering was a free-to-start web application (with a later paid tier) that generated short video clips from text prompts, animated still images, and restyled or "repainted" existing footage. The underlying models advanced quickly through several versions over 2024.[2][8]
At its March 2024 launch, Haiper's first model produced short clips, described as roughly two-second HD video and slightly lower-quality video of up to about four seconds, from text prompts or uploaded images. Features included image animation and video style repainting, with video-extension capabilities planned. The tool was offered free of charge; Miao said it was "too early" for subscription pricing in video generation and that the company wanted to "keep these features free in order to build a community." The company reported having onboarded more than 1.5 million users around this time.[1][2][8]
Released in July 2024, Haiper 1.5 roughly doubled clip length, letting users generate clips of up to about eight seconds from text, image, or video prompts. The update added an integrated upscaler that could enhance generations (and user-supplied images and videos) to 1080p in a single step, and introduced a dedicated image model so users could generate a still from text and then animate it. The longer generations and upscaler were tied to a paid Pro plan priced at $24 per month when billed annually.[9]
Haiper's flagship 2.0 model was unveiled in late October 2024 (press coverage appeared October 21, 2024, with the company promoting general availability that month). Haiper described it as a Diffusion Transformer (DiT) "perceptual" video model, combining transformer-based architectures with diffusion techniques to improve realism, temporal coherence, and generation speed. The model produced clips on the order of six seconds, could extend existing clips by about two seconds, and the company said higher-resolution output (up to 3840 by 2160, or 4K) was coming. Alongside the model, Haiper launched Video Templates, preset workflows for trending and commercial formats such as logo and product animations, face-swaps, and viral social clips. By this point the company said it had surpassed 4.5 million users, reached within roughly seven months of launch, and it described exploring video generations of up to about 30 seconds using NVIDIA technologies.[10][11][12]
In December 2024, Haiper released version 2.5, which emphasized API integrations and interoperability with third-party platforms such as the online video editor VEED, extending the model's reach beyond the consumer web app. This was the last major model version released under the independent company.[13]
Haiper's pipeline took a natural-language prompt, a reference image, or a source clip and synthesized a short video using its proprietary generative models. The company's stated technical thesis was that high-quality, controllable video generation requires a model that internalizes how the physical world behaves, including light, motion, texture, and the interactions between objects, rather than merely stitching frames together. Haiper called this direction a visual or "perceptual" foundation model and tied it to its broader goal of "emergent world modeling," the idea that a sufficiently capable video model would implicitly learn a model of the world.[2][5]
From version 2.0 onward, the company publicly characterized its system as a Diffusion Transformer, an architecture that pairs the iterative denoising of diffusion models with the scalability of transformers. Haiper said the combination improved both the realism of individual frames and the temporal consistency of motion across a clip.[10][11]
Despite rapid model iteration and a reported multimillion-user base, Haiper's consumer product was wound down less than a year after launch.
The first public signs appeared around February 3, 2025, when users posted that the Haiper web app was returning 404 errors and that they could no longer access the consumer video generator. In March 2025, both co-founders, Yishu Miao and Ziyu Wang, along with a senior machine-learning researcher (reported as Edward Hayes), updated their profiles to reflect that they had joined Microsoft AI as members of technical staff, reporting to Microsoft VP of AI Nando de Freitas. Miao wrote that after "3 years startup life with Haiper" he was "starting a new journey at Microsoft AI building multimodal generative AI."[3][4][14]
After the founders' departure, Haiper's King's Cross office closed and the roughly 20 remaining staff attempted to pivot the business toward a B2B / enterprise model rather than rebuild the consumer app. That effort did not result in an independent relaunch. In June 2025, Sifted reported that the London-based decentralized-compute company NetMind.AI had acquired Haiper's underlying video-generation models for an undisclosed sum and planned to recruit several former Haiper staff; Haiper's former VP of business, Stacie Chan, joined NetMind in a similar role. NetMind subsequently made a Haiper Text2Video model available for inference on its NetMind Power platform, repositioning the technology for API and B2B use. Commentators summarized the outcome as Haiper being "sold for parts" after Microsoft poached its founders.[4][14][15]
As of 2026, the original consumer Haiper app at haiper.ai is no longer operational as it existed in 2024, and users who stored projects only in Haiper's cloud lost access when the service was discontinued. Reporting and review sites caution that the original consumer experience has ended and that any service claiming to be the same product should be treated with skepticism.[14][16]
During its brief run, Haiper drew attention as one of several startups, alongside Runway and others, racing to commercialize text-to-video generation in the wake of OpenAI's Sora demonstrations in early 2024. Its DeepMind pedigree, fast user growth, and free consumer tier earned coverage in outlets including TechCrunch, VentureBeat, and SiliconANGLE, and reviewers tested its models against rivals such as Runway and Sora.[1][2][9][11]
Its trajectory was also read as a cautionary tale about the economics of consumer video generation. Independent analyses noted that serving free or low-priced video generation to millions of users is computationally expensive, and that well-capitalized incumbents and the high cost of GPUs make it difficult for a seed-stage startup to sustain a standalone consumer product. The combination of competitive pressure, capital intensity, and a high-profile talent acquisition by Microsoft was widely cited as the backdrop to Haiper's wind-down.[3][4][14]