Paradigm Intelligence, formerly 4Paradigm, has unveiled its in-house models for the first time, launching two at once. Phai-A1 is a general-purpose language model focused on coding and AI assistants, while Phai-M1 generates video for live-action film and television. Both are available to try at phai.phanthy.com.

The company highlights two core figures for A1: at comparable quality, it responds eight to 10 times faster than models in the same parameter class; and it can run on a single consumer GPU while supporting a native context window of more than 1 million tokens. For M1, the pitch is simpler: about three minutes to generate a 15-second 720p video with synchronized audio and visuals.

The 8-10x Speedup Was Measured on a Single RTX 3090

The company spells out how it measured that multiple: both models ran on a single RTX 3090, using the same version of an unnamed coding agent to perform identical tasks. The comparison measured relative speed on the same GPU, framework and workload.

The more specific the conditions, the more credible the result, but also the less broadly it can be applied. Change the task, agent framework or GPU, and the multiple no longer necessarily holds. The company also qualifies the claim with “while maintaining quality,” and it is the company itself that decides whether the quality bar has been met.

More interesting is the company's assertion that “these figures were not produced under ideal laboratory conditions.” In fact, the opposite is true: precisely because the GPU, framework and task were all fixed, this is a result for one specific scenario, not a measure of general-purpose throughput.

The evaluation involved three benchmarks: SWE-Bench Verified, SWE-Bench Pro and SWE-Bench Multilingual. The company says A1 outperformed peers in its class and even matched some frontier models with hundreds of billions of parameters. But it did not disclose a single score for any of the three benchmarks.

The phrase “matched some” is also worth parsing. The emphasis is on “matched,” while “some” narrows the scope. Given that the models differ in parameter count by dozens of times, the wording leaves more room than it may first appear.

There is also a set of user ratings. A1 spent a period running anonymously on PhanthyCode and ranked first among all integrated open-source and commercial models in three categories: model capability, stability and speed, and overall user satisfaction. But this was a user vote within the company's own product, and it disclosed neither the number of participants nor the voting period or sample-selection methodology.

Ascend's Role Stops at Post-Training

The section on Chinese computing power is the part of the article that deserves the closest scrutiny.

The introduction says the models were trained end to end on Huawei Ascend chips. But in the detailed section below, the wording changes: the Phai series underwent specialized post-training optimization across the full workflow on Huawei's Ascend platform.

Training and post-training are not the same. Pretraining feeds a model trillions of tokens from scratch and can cost tens or even hundreds of millions of dollars. Post-training applies supervised fine-tuning and reinforcement learning to an already trained base model, at a cost one or two orders of magnitude lower. The former determines whether a model can speak; the latter determines how well it speaks.

The next sentence says the models were “not trained first and adapted later,” stressing that they were designed for a Chinese computing environment from the start of the model-optimization stage. That distinction matters: adaptation means porting an existing model and getting it to run, while post-training means tuning it again on Chinese chips. The two are a level apart. But this proves that post-training took place on Ascend, not that pretraining did.

Both claims appear in the same article: the summary says “training,” while the body says “post-training.” Readers usually remember the summary.

The investment in China's domestic ecosystem is tangible. HAMi manages heterogeneous computing power, while the Xinchuang Model Box community handles compatibility between chips and models. The article also notes reports from two months ago that Paradigm spent more than 1 billion yuan on Huawei Ascend 950 chips. That was attributed to unnamed reports, not a company announcement.

A Million-Token Context Window and a Laptop

Another selling point for A1 is its long context window. The company emphasizes that it is “native”: the post-training data directly covers scenarios involving more than 1 million tokens, unlike the “stacked” approaches to long context used elsewhere in the market.

The two approaches differ in cost and stability. A stacked approach extends the window after training through techniques such as positional-encoding extrapolation or interpolation. It avoids that expense during training, but the expanded window is effectively borrowed, and long-range retrieval tends to degrade. A native approach feeds long sequences directly into post-training. It is more stable, but because attention computation rises with sequence length, training is far more expensive. Choosing a native approach means paying that bill during training.

The company says this allows A1 to navigate large code repositories, lengthy documents and conversations spanning several days. But it provides no benchmark results to support the claim, whether recall rates at 1 million tokens or accuracy on long-document question answering.

Then there is the claim that “an ordinary laptop can run it with a double-click.” Two figures are missing: the precision level after quantization and the amount of memory required for the cache at a 1-million-token context length. The longer the context, the more memory the KV cache consumes; that cost cannot be avoided. The article mentions both a “single RTX 3090” and an “ordinary laptop,” but gives no memory figures for either.

Recalculating M1's Economics

The section on M1 is shorter. Positioned for live-action film and television, the model received specialized training in character performance, visual detail and overall production quality, which the company says has reached cinematic standards. Its main selling point is synchronized audio and visuals: sound and imagery are created within the same generation pipeline, eliminating the need for lip-syncing in post-production.

It takes three minutes to generate 15 seconds of 720p video. The article makes its own economic case: for AI short-drama studios, generation time and speed directly determine production cadence and cost. That argument holds. Short dramas are budgeted by episode and minute, so the actual throughput of a production line is the generation time per clip multiplied by the number of clips and then by the number of retries.

But 15 seconds is a very small unit. A two-minute episode requires eight 15-second clips. If each usable clip takes several attempts, the calculation is three minutes multiplied by eight and then by the number of draws. The article does not finish that calculation. It gives 720p as the resolution but says nothing about generation speeds at higher resolutions.

Open Source Is Promised “Soon”

The model weights are not yet available. The article uses “soon” twice: the weights will soon be released on Hugging Face and ModelScope; and the models will soon be open-sourced, alongside more detailed technical reports.

Until those technical reports are published, the methodology behind the eight-to-10-times speedup, the scores on the three benchmarks and real-world performance at a 1-million-token context length all remain based on the company's own testing. For now, the only thing users can verify is the web demo.

The models have already been fully integrated into the company's own products: PhanRouter, ModelSphere, Phanthy, PhanthyMovie and PhanthyCode/Work.

When a company with more than a decade of enterprise deployment experience releases a foundation model for the first time, the key question is not how fast it runs. It is which stages have moved onto Chinese chips and which remain elsewhere. The answer provided here is clear: post-training took place on Ascend, while nothing was said about pretraining.