Slop TVNewsLatest
News

Phancy's Phai-M1 generates video and audio in one pass

Phancy says it plans to release the model's weights, but only Phai-A1's are downloadable today; Phai-M1 has an online trial.

Illustration: Phancy's Phai-M1 generates video and audio in one pass
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • Phancy Group announced Phai-M1 on November 18, 2026, a video model that generates live-action-style footage and its dialogue, sound effects and ambience in one pass.
  • Phancy says Phai-M1 renders 15 seconds of 720p video in about three minutes on a single consumer GPU with 16 GB of memory, a company test that has not been independently reproduced.
  • Phancy says it plans to release Phai-M1's weights, but only Phai-A1's are downloadable now, on Hugging Face and ModelScope; Phai-M1 has an online trial and its weight download is listed as coming soon.
  • Phancy's Phai-A1 is a post-trained derivative of Qwen3.6-35B-A3B, not a model trained from scratch, according to the model card and Pandaily.

Phancy Group announced Phai-M1, a video model that generates live-action-style footage and its audio in one pass.

The company announced the model on November 18, 2026, alongside a second in-house model, Phai-A1, in a release published through Media OutReach Newswire and mirrored on the maker's own page at phai.phanthy.com.

Phai-M1 turns text or an image into video with sound, at clip lengths from 5 to 20 seconds and resolutions up to 1080p, in aspect ratios including 16:9, 9:16, 4:3, 3:4 and 1:1. The maker's page says dialogue, sound effects and ambient audio are generated with the picture rather than dubbed in after. It lists four generation modes: text to video, image to video, first and last frame, and reference generation, which steers faces, setting and sound from a subject image, a background image and a reference audio clip.

The speed figure is a company number. Phancy's page says Phai-M1 renders 15 seconds of 720p video in about three minutes on a single consumer GPU with 16 GB of memory. The company ran that test at 720p and has published no independent results.

Phancy is the enterprise AI cloud company formerly known as 4Paradigm, listed in Hong Kong as 6682.HK. Phai is its first in-house model family. Phancy says the series was post-trained end to end on Huawei's Ascend platform with support from the ModelHub XC community, and it is wired into the company's ModelHub XC, HAMi and ModelSphere products and its PhanthyCode and PhanthyMovie apps. Its founder and chief executive, Dr. Dai Wenyuan, said in the release that the industry is moving "from competition over model capabilities to a phase in which computing infrastructure, models and applications are developed together."

Phai-A1 is the other half of the launch. The model card identifies its base as Qwen3.6-35B-A3B, a 35-billion-parameter mixture of experts, so it is a post-trained derivative rather than a model trained from scratch. Phancy reports 83.6% on SWE-bench Verified and says Phai-A1 answered about 8 to 10 times as fast as Qwen 27B on identical coding-agent tasks in its own tests.

The caveat is the weights. Phancy says it plans to make the weights for Phai-A1 and Phai-M1 available to developers "in the near future." Only one is downloadable now. Phai-A1's weights are on Hugging Face and ModelScope; Phai-M1 has an online trial, and the maker's page lists its weight download as coming soon. Pandaily reports the A1 repository metadata lists Apache 2.0 with the underlying Qwen license also applying, across 16 BF16 shards totaling roughly 70 GB.

For a creator, the pitch is the single-pass audio. Removing the separate dub-and-lip-sync step is where AI short-drama pipelines lose time, if the synchronization holds up outside the company's own clips.

Try Phai-M1 in the browser at phai.phanthy.com/console.

Sources

  1. phai.phanthy.com - the maker's page: M1 specs, four modes, 15s/720p/3min claim, weights "coming soon", console
  2. media-outreach.com - the announcement release: launch date, M1 description, planned weights, Ascend, integrations
  3. huggingface.co - A1 model card: Qwen3.6-35B-A3B base, Ascend post-training, GRPO 4,000 sandboxes, 16K-1M tokens
  4. pandaily.com - the A1-downloadable vs M1-trial distinction, repo inspection, Apache 2.0, ~70GB across 16 shards
  5. qbitai.com - independent Chinese-language coverage: Ascend end-to-end training, the 8-10x and 3-min/15s claims
  6. channelnewsasia.com - republication of the same Media OutReach release (one source)
  7. markets.ft.com - the company announcement (same release)