Modelsslug: phai-a1

Phai-A1

Phanthy AI's general-purpose language model for coding and agent tasks, runnable on a single GPU with native million-token conte

Curated Intelligence

3 items · Page 1/1
InfoQ 中文T2·Models
72 pts
Paradigm Unveils Phai Models Trained on Domestic Compute with 10x Speedup

Paradigm Intelligence unexpectedly released the Phai-A1 language model and Phai-M1 video model on November 18. Phai-A1 uses a MoE architecture with 35B total parameters but only 3B activated per token, achieving 8-10x faster response than models of similar scale at equal quality, and natively supports million-token context. Both models are trained on domestic compute, marking a shift from inference to large-scale training for China's AI compute ecosystem.

量子位官网T2·Models
72 pts
Phanthy Unveils Phai-A1 and Phai-M1 Trained Entirely on Ascend Chips

Phanthy (formerly 4Paradigm) has unveiled its first in-house foundation models: Phai-A1, a general language model, and Phai-M1, a video generation model for live-action film scenarios. Both were trained end-to-end on Huawei Ascend compute. Phai-A1 targets coding and agent workloads, claiming 8-10x faster responses than peers of similar size while running on a single RTX 3090 with native million-token context; Phai-M1 produces 15-second 720P clips with synchronized audio in about three minutes. Weights are slated for release on Hugging Face and ModelScope.

量子位 (微信公众号)T2·Models
72 pts
Phanthy AI Unveils Phai-A1 and Phai-M1 Trained Entirely on Ascend Chips

Phanthy AI (formerly 4Paradigm) has unveiled its first self-developed foundation models: Phai-A1, a general-purpose language model, and Phai-M1, a video generation model for live-action film scenarios. Both were trained end-to-end on Huawei Ascend chips. Phai-A1 targets coding and agent workflows, claiming 8-10x faster responses than same-scale models at equal quality, running on a single RTX 3090 with native million-token context; Phai-M1 produces 15-second 720P video with synchronized audio in about 3 minutes. Weights are slated for release on Hugging Face and ModelScope.