Paradigm Intelligence launches Phai-A1 language model and Phai-M1 video model, trained on Huawei Ascend
Paradigm Intelligence (范式智能, formerly 第四范式) has unveiled two in-house models: Phai-A1, a language model that runs on a single RTX 3090, and Phai-M1, a video model with synchronized audio. The company says both were trained on Huawei Ascend hardware, with open weights coming soon.
Paradigm Intelligence (范式智能, formerly 第四范式) has unveiled two in-house foundation models, Phai-A1 and Phai-M1, and says both were trained across the full pipeline on Huawei Ascend hardware. The announcement was reported on 2026-11-19 by QbitAI (量子位), a Chinese-language outlet. This is the company's first time putting its own models front and center.
The company says weights for both models will be released on Hugging Face and ModelScope soon, along with detailed technical reports. Pricing, parameter counts and training cutoff dates are not yet disclosed.
Phai-A1: general language model for coding and assistants
Phai-A1 is a general-purpose language model aimed at code development and intelligent assistant work: writing code, fixing bugs, calling tools and running multi-step tasks. According to the company:
- Speed: At comparable quality, response speed is about 8-10 times faster than models of the same parameter class.
- Hardware: It runs on a single consumer GPU. In the company's comparison, both models ran on one RTX 3090 with the same version of an unnamed code agent performing identical tasks.
- Context: It has native million-level context. The company says post-training data directly covers million-token scenarios, as opposed to stacked or extended long-context approaches.
- Benchmarks: On SWE-Bench-Verified, SWE-Bench-Pro and SWE-Bench-Multilingual, Phai-A1 outperforms models of its class and is claimed to be not weaker than some frontier models at the hundred-billion-parameter scale. Specific scores were not given.
The company also says Phai-A1 went through an anonymous-model period inside its AI office product PhanthyCode. Among all connected open-source and commercial models, it reportedly received the highest ratings on model capability, stability and speed, and user satisfaction. These are company-reported results.
Phai-M1: video generation for live-action film and TV
Phai-M1 is a video generation model trained specifically for live-action film and TV scenarios. The company says it reaches cinematic quality in character performance, visual detail and overall texture. Audio and video are generated in the same pipeline, so lip-syncing does not need to be done in post-production. Generation takes about 3 minutes for a 15-second 720P video.
Trained on Ascend
Paradigm says the Phai series completed specialized post-training optimization on the Huawei Ascend platform, with support from the ModelHub XC community. The company says the work was designed around domestic compute from the start of model optimization, rather than adapted after training. QbitAI states that few Chinese AI companies can currently do full-chain training natively on domestic compute.
QbitAI also points to Paradigm's earlier ecosystem work, including the HAMi heterogeneous compute management project and a chip-model adaptation community. It adds that two months ago there were reports that Paradigm would purchase Huawei Ascend 950 hardware worth over 1 billion yuan.
Availability
The Phai series is already integrated into Paradigm products: PhanRouter, ModelSphere, Phanthy, PhanthyMovie and PhanthyCode/Work. A trial portal is available at phai.phanthy.com.
What this means
Phai-A1's central claim is a practical one: coding-agent performance on a single RTX 3090, without cloud queues or API billing. The 8-10x speed figure comes from the company's own comparison, so it will need independent checks once weights are public, as will the SWE-Bench results, for which no scores have been published yet.
For Phai-M1, the stated 3 minutes for a 15-second 720P clip matters mainly for production teams, where generation time affects cost and pacing. Training both models on Ascend, if confirmed by the technical reports, would show that the platform can support a language model and a video model in the same release. The open-weight release will determine how far outside developers can test any of this.
Related Articles
Apple releases FastFrames-L, a 14B two-step video model distilled from Wan2.1-14B for interactive 480p generation
Apple has released FastFrames-L (14B) and FastFrames-S (1.3B), autoregressive two-step video diffusion models that generate 480p video at interactive latency. FastFrames-L is distilled from Wan2.1-14B, and the repository lists 35.4 ms total latency for FastFrames-S on one GPU.
Cloudflare releases open-weight Clairic 27B and new Clairic-flash 9B for data extraction
Cloudflare announced on November 18, 2026 an updated 27-billion-parameter Clairic and a new 9-billion-parameter Clairic-flash. Both have open weights on Hugging Face and are available as hosted Workers AI endpoints.
RadixArk releases DaisyRose-27B, lifting Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%
RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning on 1,200 curated agent tasks. According to RadixArk, it raises Terminal Bench 4.1 pass@1 from 5.6% to 12.0% and scores higher than the base model on all seven reported benchmarks. Weights ship under Apache 2.0 in BF16, FP8 and NVFP4.
Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license
Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.
Comments
Loading...