Phancy Launches Phai-A1, a 1M-Token Model That Runs on a PC, and Phai-M1 for AI Video With Sound

Hong Kong-listed Phancy launched Phai-A1, a fast language model with a one-million-token context that runs on a PC, and Phai-M1, a video model for live-action-style footage with synchronised audio. Both were tuned for Huawei Ascend and weights are planned.

Christopher Ort
By Christopher Ort · 3 min read
Share
A cinema camera on a dolly beside an empty director's chair on a film set.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. How the models work
  5. Why it matters
  6. What's still open
  7. Sources
  8. i10x Perspective

Hong Kong-listed Phancy Group has launched two in-house AI models: Phai-A1, a language model with a native one-million-token context window that can run on a personal computer, and Phai-M1, a video generation model built for live-action-style footage with synchronised audio. The company announced both on 18 November 2026 and says it plans to release the weights for both models to developers worldwide “in the near future.”

i10X Quick Take

Quick Take

  1. Phai-A1: general-purpose language model for coding, bug fixing, tool use and multistep tasks, with a native context window of about one million tokens and local deployment on a PC.
  2. Speed claim: in Phancy’s single-GPU tests on identical coding-agent tasks, Phai-A1 responded about 8 to 10 times as fast as Qwen 27B.
  3. Phai-M1: generates video and synchronised audio in one workflow; a 15-second 720p clip took about three minutes in testing.
  4. Hardware: both models were post-trained and optimised for Huawei’s Ascend computing platform.
  5. Weights: planned open release, no date given. The models can be tried on Phancy’s Phai web platform now.

In plain English

Phancy, a Chinese enterprise AI cloud company, has built its own text model and its own video model. The text model is small and fast enough to run on a strong personal computer, yet can read a huge amount of material at once, such as a whole code project. The video model is aimed at studios making live-action-style series and short dramas, and it creates the sound together with the picture. Phancy says it will publish both models for anyone to download later.

What happened

Phancy (stock code 6682.HK) announced the Phai series through Media OutReach Newswire from Beijing and Hong Kong. Founder and CEO Dr. Dai Wenyuan said: “The AI industry is moving from competition over model capabilities to a phase in which computing infrastructure, models and applications are developed together. The Phai series is an important step in strengthening our full-stack AI capabilities.”

The independent LLM Releases tracker lists the language model as Phai-A1-35B-1M, released on 18 November.

How the models work

  • Phai-A1 benchmarks: Phancy says it performed well on SWE-bench Verified, SWE-bench Pro and SWE-bench Multilingual and was comparable to larger models on some evaluations, but the release does not list scores.
  • Long context: post-training covered very long texts so the model can handle whole repositories, long documents and extended multi-turn work.
  • Phai-M1 focus: trained on live-action film and TV scenes to render facial expressions and body movement more naturally and reduce common distortions in AI video.
  • Product stack: both models are integrated with Phancy’s ModelHub XC, HAMi and ModelSphere infrastructure products and its PhanthyCode and PhanthyMovie apps.

Why it matters

For developers, a fast long-context coding model that runs locally is useful for private codebases and for agent workloads that make many model calls, once the weights are out.

For video teams, Phai-M1 joins a crowded field of Chinese video models, alongside ByteDance’s Seedance, aimed at the fast-growing market for AI-generated short dramas.

For the chip market, Phancy is another Chinese AI company tuning its models for Huawei Ascend hardware rather than Nvidia GPUs.

What’s still open

  • Published benchmark scores, parameter counts and licence terms from Phancy itself.
  • The release date for the model weights.
  • Pricing for API or enterprise use.

Sources

  • Phancy Group, “Phancy Launches Phai-A1 Language Model and Phai-M1 Video Generation Model,” Media OutReach Newswire, 18 November 2026
  • LLM Releases tracker, entry for Phai-A1-35B-1M, 18 November 2026

i10x Perspective

The Phai launch shows how Chinese AI cloud providers are building their own models to sell more compute and apps, and tuning them for domestic chips. The claims are company-reported until the weights land and independent tests follow. For context, see Seedance 2.5 4K, ByteDance’s latest video model and more launches in our Models section.

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Editor's picks

  1. GAO Report: ChatGPT, Claude, Gemini and Grok Produced Viable Attack Plans After Safeguards Were Bypassed
  2. ResolveAI SRE-1: A 27B Model Trained to Find the Root Cause of Production Incidents
  3. Reson8 Rumble-2-Flash: Open-Weight 0.5B Speech Recognition Model Covers 40+ Languages
  4. Pika Enact Lets Creators Direct AI Video Actors and Cameras in Real Time
  5. Google Stitch Loop: An Always-On Product Agent That Fixes Security, Speed and Accessibility via Pull Requests