Phancy Group has launched two AI models under its new Phai family, a coding model built for long-context agent work and a video model that generates synchronized audio. The Hong Kong listed enterprise AI company announced Phai-A1 and Phai-M1 on November 18, 2026, pitching the pair as the model layer of its full stack of compute, platform and application products.
Phai-A1 is a general-purpose language model tuned for software development and AI assistants, with code generation, bug fixing, tool use and multistep tasks as its core jobs. Phancy says it natively supports a one-million-token context window, long enough to hold entire code repositories, long documents and extended multistep conversations in a single session. The model can also run locally on a personal computer, depending on hardware, which the company frames as a flexibility and data control option for developers and enterprises.

What Phai-A1 Claims on Benchmarks
On software engineering tests, Phancy reports that Phai-A1 scores 83.6% on SWE-Bench Verified and 65.0% on SWE-Bench Pro, with 74.0% on SWE-Bench Multilingual. Its official benchmark chart puts those results close to larger rivals, including GLM-5.2 at 82.8% on SWE-Bench Verified and Qwen3.8-27B at 86.0%, while Phai-A1 leads its own base family. The company also claims the model answered coding agent tasks 8 to 10 times as fast as Qwen 27B under equivalent single-GPU conditions. These are company reported figures, with no independent verification published yet.
The official Phai site describes the model as a 35B mixture-of-experts design, and its Hugging Face listing names Alibaba’s Qwen3.6-35B-A3B as the base model, with Phancy’s own post-training on top. That post-training ran end to end on Huawei’s Ascend computing platform, according to the company, and the model is integrated with Phancy’s ModelHub XC, HAMi and ModelSphere products, plus applications including PhanthyCode and PhanthyMovie. The weights are listed for download on Hugging Face and ModelScope under the Apache 2.0 license.

Phai-M1 Targets Live-Action Video
Phai-M1 is a video generation model trained for live-action film and television scenes. Phancy says it renders facial expressions and body movements more naturally, improves visual detail and reduces common distortions in AI-generated video, aiming at live-action series and AI-generated short-form dramas. Unlike many video models, it generates visuals and synchronized audio in a single workflow. In company testing, it produced a 15-second video at 720p resolution in about three minutes. Phancy offers a trial of the Phai models at phai.phanthy.com.
Why It Matters
Phancy’s pitch is vertical integration rather than raw scale. Founder and chief executive Dr. Dai Wenyuan said the industry is moving from competition over model capabilities to “a phase in which computing infrastructure, models and applications are developed together”, adding: “By optimizing our in-house models and computing platforms together, we can connect the underlying infrastructure with business applications and make our models better suited to practical business needs.”
That framing matters because Phai-A1 arrives as a practical, openly licensed option rather than a frontier challenger: a 35B model with a million-token window, local deployment and an Apache 2.0 license is aimed squarely at developers and enterprises that want to run coding agents on their own hardware. The open question is how the company reported benchmark and speed claims hold up once independent testers run the published weights, which the company says will be available to developers worldwide.
Sources
- Phancy Launches Phai-A1 Language Model and Phai-M1 Video Generation Model, Media OutReach Newswire via Malay Mail, November 18, 2026
- Phai official site, Phancy
- Phai-A1-35B-1M model page, Hugging Face

