An editor opens a clip of a shop assistant speaking to camera. The mouth moves with the words, the hands follow the sentence, and the room tone stays steady when the voice pauses. Nothing was dubbed on afterwards.
That is the promise behind a new video model from a company better known for selling software to banks.
What Was Announced
On 18 November 2026, in Beijing, Paradigm Intelligence Technology Group (范式智能技术集团股份有限公司, 6682.HK) — formerly 4Paradigm (第四范式) — released two models under the Phai name at its twelfth-anniversary event.
- Phai-A1 is a language model aimed at code development and assistant tasks. The company says it responds roughly eight to ten times faster than Qwen 27B on a single RTX 3090 graphics card, supports a native one-million-token context window, and runs on a personal computer.
- Phai-M1 is a video generation model (视频生成模型) trained for live-action film and television scenes. The company says it generates audio and picture in the same pass, takes about three minutes to produce a 15-second clip at 720p, and improves the naturalness of facial expressions and limb movement.
The company also named the applications already connected to the models: PhanRouter, ModelSphere, Phanthy, PhanthyMovie and PhanthyCode. Founder and chief executive Dai Wenyuan (戴文渊), who started the company in 2014, described the launch as a step from competing on single-model capability toward coordinating computing power, models and applications.
Both models were optimised through training and post-training on Huawei's Ascend (昇腾) platform. That detail is doing a lot of work in the announcement. For Chinese enterprise buyers, a model that can be adapted on domestic hardware is increasingly a procurement requirement rather than a talking point.
Why Audio in the Same Pass Matters
Most AI-produced video still separates the two jobs. A model generates silent pictures; a second system creates a voice or a music bed; an editor lines them up. The seams show. Lips drift, background noise changes between shots, and a character's breath no longer matches the pause in speech.
Generating sound and image together removes one source of that mismatch. The model has to decide what the scene sounds like while it decides what the scene looks like, which is harder but more coherent. Phai-M1's stated use cases — digital humans (数字人), short clips and advertising material — are exactly where a visible mismatch costs the most. A fifteen-second advertisement with lip-sync that slips is not a technical curiosity; it is unusable.
The three-minute figure deserves a caveat. It describes a 720p clip under the company's own test conditions. It says nothing about how many attempts were needed, what happened at higher resolution, or how the model performs on a scene with several speakers. For a microdrama (短剧) team planning a production schedule, those unknowns matter more than the headline speed.
The Live-Action Problem
Chinese video models have spent 2026 competing on length, resolution and reference-image control. Phai-M1 is aiming at a different weakness: the look that practitioners call "plastic". Faces warp, hands lose fingers, and every subject seems lit by the same soft studio light.
The company says it trained specifically on real-person footage and optimised for expression and body movement. It is a sensible target, because live-action material is where video models meet paying customers. Advertising agencies and short-drama studios care less about a model's benchmark score than about whether a client will accept the footage.
But "less plastic" is a qualitative claim. The announcement published no blind comparison, no user study and no side-by-side evaluation against competing models. Until third parties test it, the honest description is that Paradigm Intelligence is targeting realism, not that it has achieved it.
The Open-Weights Question
Here is the part that needs careful wording. Some coverage of the launch described Phai-A1 and Phai-M1 as "released and open-sourced". The company's own statement, as reported in the technology press, says it plans to open the model weights on Hugging Face and ModelScope in the near future.
Those are not the same thing. A launch with hosted access, or access through a company's own applications, lets you try a system. A launch with weights lets you inspect, benchmark, fine-tune and run it locally. The distinction determines whether a claim can be independently checked.
At the time of writing I could find no published Phai model weights. This article therefore treats the performance figures as company statements. If and when the weights appear, those figures become testable.
Where This Fits
Phai-M1 arrives into a crowded field. Kling, Seedance, Wan and Vidu have all released video models in 2026, several with longer clips or higher resolution. Competing on those terms alone would be difficult.
Paradigm Intelligence's angle is integration. The company sells enterprise software, and the Phai models are already wired into its own product suite. For a bank or a retailer, a video model that sits inside an existing workflow may be more valuable than a leaderboard winner that does not. That is a distribution argument, not a technical one — and distribution is often what decides which model a company actually uses.
Key takeaways
- Paradigm Intelligence released Phai-A1 (language) and Phai-M1 (video) in Beijing on 18 November 2026.
- Phai-M1 generates synchronized audio and video in one pass; the company cites roughly three minutes for a 15-second 720p clip.
- Both models were optimised on Huawei Ascend (昇腾) domestic computing infrastructure.
- The models are connected to the company's own applications, but the promised Hugging Face and ModelScope weights were not available at launch.
- No independent evaluation of the video quality or speed claims has been published.
Honest limitations
Every performance figure in this piece — the eight-to-tenfold speed-up, the one-million-token context, the three-minute generation time — is the company's own statement. I found no independent benchmark for either model.
There is a source discrepancy on the open-weights question. Some headlines called the release open-source; the company's statement, as reported by technology media, promised weights "in the near future". I have followed the company's wording. If weights have since appeared, the situation has changed and readers should verify the current state.
I have not tested Phai-M1, and I could not confirm whether a public demo exists outside the company's own applications. No pricing was published in the material I reviewed, so a cost-per-clip comparison is not possible.
No currency amounts appear in the launch material I used, so no US$/HK$ conversion applies to this article.
What readers can do now
- Check whether weights exist before treating a benchmark as verified. "Coming soon" and "available" are different states, and only one of them allows independent testing.
- For video work, test lip-sync on your own footage. A model that looks impressive on a studio demo can fail on a product shot with side lighting and two speakers.
- Ask what a model is connected to, not just what it can do. A mediocre model inside a working production workflow often beats a better one that requires a new pipeline.
