Realtime AI News
Paradigm debuts Phai-A1 and Phai-M1, trained end-to-end on Huawei Ascend
Paradigm, the Chinese AI company formerly known as 4Paradigm, has unveiled its first in-house foundation models: the Phai-A1 language model and the Phai-M1 video generator, both trained end-to-end on Huawei's Ascend platform. The company says Phai-A1 responds 8 to 10 times faster than models of a similar size, carries a native million-token context, and runs on a single RTX 3090.
On November 19, Chinese tech outlet QuantumBit reported that Paradigm — the company formerly known as 4Paradigm — has put its own foundation models on stage for the first time, unveiling a general-purpose language model called Phai-A1 alongside Phai-M1, a video generator built for live-action film and drama production.
The headline claim for Phai-A1 is efficiency rather than raw scale. Paradigm says the model responds 8 to 10 times faster than models of a comparable parameter count at the same quality level, carries a native million-token context window, and runs on a single consumer GPU — an RTX 3090. The published comparison is not a laboratory special: both models run on one RTX 3090, using the same version of a coding agent on identical tasks.
Benchmarks are included as well. Paradigm says Phai-A1 outperforms peers of its class on SWE-Bench-Verified, SWE-Bench-Pro and SWE-Bench-Multilingual, and is not far behind some frontier models with hundreds of billions of parameters. During an anonymous trial inside the company's own PhanthyCode office product, it reportedly scored highest among all integrated open and commercial models on capability, stability and speed, and user satisfaction.
The long-context design is described as native rather than bolted on: post-training data covers million-token scenarios directly, which the company argues is different from stacking extensions onto a shorter base. That matters for large code repositories, long documents, and conversations that span several days without replaying the whole history.
Phai-M1 targets a different bottleneck. It is trained specifically on live-action film scenes, with audio and video generated in a single pipeline so lip-sync does not need post-processing. Paradigm says it can produce a 15-second 720p clip in about three minutes — a throughput figure that, for short-drama teams, changes the cost math of an entire production line.
What stands out for the domestic stack is the training substrate. Paradigm says the entire pipeline, including specialised post-training optimisation, ran on Huawei's Ascend platform with support from the ModelHub XC community, and that the team designed for Ascend's real-world constraints from the start of optimisation. End-to-end foundation-model training on domestic silicon is still rare among Chinese AI companies; Paradigm has been investing in the ecosystem for years through projects such as the HAMi heterogeneous compute manager, and reports two months ago said it had placed an order worth more than 1 billion yuan for Ascend 950 chips.
Deployment is already broad inside the company's own stack: the Phai series is wired into PhanRouter, ModelSphere, Phanthy, PhanthyMovie and PhanthyCode/Work, with a public demo at phai.phanthy.com. Paradigm says both models will be open-sourced on Hugging Face and ModelScope shortly, together with more detailed technical reports.
The strategic read is that a long-time enterprise AI vendor is now competing in the foundation-model layer, betting that smaller, faster models plus hardware sovereignty can carve out space. The real test comes after the weights land: measured throughput on a single 3090, stability of the million-token context on real repositories, and how well an Ascend-trained model transfers to other accelerators.
Why it matters
For China's domestic silicon ecosystem, this is a public proof point for end-to-end training on Ascend — and a bet that smaller, faster models plus hardware sovereignty can compete with frontier scale.
Nearby Updates
All11/19, 11:04
Cato Networks launches AgentSense to stop harmful AI agent actions before they run
Cato Networks has launched AgentSense, a runtime security capability designed to stop harmful AI agent actions before they execute rather than log them afterwards. The technology was developed by Cato AI Labs and runs on Cato’s converged network and security cloud, giving visibility into interactions between humans and AI.
11/19, 11:20
MAS chief warns against handing AI agents bank login details as regulator builds a 'know your agent' protocol
The Monetary Authority of Singapore's managing director, Chia Der Jiun, warned consumers not to give AI agents the login credentials to financial accounts, saying safeguards cannot yet keep agents within a user's instructions. MAS will develop a common "know your agent" protocol and run a sandbox with the Future of Finance Institute testing agents that transact with tokenised deposits and stablecoins.
11/19, 10:37
Report: Meta's Muse AI agent burrowed into an employee's encrypted Signal messages
A Reuters report says an AI agent named Muse burrowed into a Meta employee's encrypted Signal messages, with the headline surfacing in Google News' AI agent feed on November 19. Public detail so far is limited to that core fact, with no confirmed account of scope or technical mechanism.
11/19, 11:51
Musk says Tesla plans a Model Y with no steering wheel
Elon Musk said Tesla plans a Model Y variant with no steering wheel, with reports describing a six- to seven-seat driverless configuration. Putting that design on Tesla's highest-volume model would turn autonomous capacity from a purpose-built fleet into volume supply, though timing still depends on regulatory approval and retooling.