- Phai-A1 uses MoE with 35B total parameters and only 3B activated per token, reducing per-token compute by an order of magnitude, the main source of speedup.
- Native million-token context is enabled via YaRN extra training and rope scaling, maintaining speed on long tasks.
- Models are trained on domestic compute (e.g., Huawei Ascend), indicating China's AI compute ecosystem is moving from inference to large-scale training.
AI CURATED HOT STORY · 11/18/2026
范式智能 Unveils Phai Models Trained Entirely on Ascend Chips
范式智能 (formerly 第四范式) unexpectedly released its self-developed foundation models Phai-A1 and Phai-M1 on November 18, with both models trained entirely on Huawei Ascend compute, marking the first validation that domestic chips can support foundation model R&D. Phai-A1 is a general-purpose language model using a MoE architecture with 35 billion total parameters but only 3 billion activated per token, delivering 8-10x faster response than models of similar scale at equal quality, runnable on a single RTX 3090, and natively supporting million-token context. Phai-M1 is a realistic video generation model with audio-visual sync, producing 15-second 720P videos in about 3 minutes. Both models will be open-weighted on Hugging Face and ModelScope.
Across sources, Phai-A1's speed advantage stems from the MoE architecture reducing per-token computation by an order of magnitude, while YaRN extra training and rope scaling enable million-token context without sacrificing responsiveness. Phai-M1 targets film-grade content generation with emphasis on audio-visual sync and efficiency. Both models trained on domestic compute signal that China's AI chip ecosystem is moving from inference to large-scale training.
The release carries multiple implications: 范式智能's first public foundation models demonstrate its transition from an AI platform to foundation model R&D; the full-stack Ascend training provides crucial validation for domestic compute, potentially encouraging more enterprises to adopt Chinese chips for large model training; and the open-weight strategy lowers barriers for developers and fosters ecosystem growth.
Hourly heat curve
30 fully observed hoursThe line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.
Coverage timeline
Total 3 reports · Latest first- InfoQ 中文T2·72 pts
- 量子位官网T2·72 pts
- Both Phai-A1 and Phai-M1 were trained end-to-end on Huawei Ascend chips, validating domestic compute for foundation model development
- Phai-A1 targets coding and agent tasks, runs on a single RTX 3090, and claims 8-10x faster responses with native million-token context
- Phai-M1 generates 15-second 720P video with synced audio in about three minutes; both models will be released as open weights
- 量子位 (微信公众号)T2·72 pts
- Both models were trained end-to-end on Huawei Ascend chips, validating domestic compute for foundation model development
- Phai-A1 targets coding and agent tasks, runs on a single RTX 3090, and natively supports million-token context
- Weights will be released on Hugging Face and ModelScope soon