Phancy’s Phai-A1 coding model is post-trained from Qwen3.6-35B-A3B
Phancy calls its coding model Phai-A1 its own, but the model card names Qwen3.6-35B-A3B as the base. The weights can already be downloaded, and the performance figures are the company’s own.
Phancy, the Hong Kong-listed company formerly known as 4Paradigm, unveiled the Phai-A1 language model and the Phai-M1 video model on 18 November. Phai-A1 is designed for coding and agentic work. In a press release distributed through Media OutReach, the company says it developed the models itself and promises developers the weights ‘in the near future’.
Pandaily reported, however, that the Phai-A1 weights are already on Hugging Face and ModelScope, and that the repository timestamps predate the announcement. The model card Pandaily read names Qwen3.6-35B-A3B, a 35-billion-parameter mixture-of-experts (MoE) model, as the base. In an MoE model, only some of the parameters process each token. Phai-A1 is therefore a post-trained derivative, not a model Phancy trained from scratch. Phancy said fine-tuning and reinforcement learning ran on Huawei’s Ascend processors. During reinforcement learning, feedback came from code execution, unit tests and tool calls in 4,000 sandboxes.
Phancy reports scores of 83.6 per cent on SWE-bench Verified, 65.0 per cent on SWE-bench Pro and 74.0 per cent on SWE-bench Multilingual. These benchmarks measure what share of tasks from real software projects a model solves. There are no independent measurements. In the company’s own chart, Qwen3.8-27B scores 86.0 per cent on SWE-bench Verified, Gptonline.ai reported.
The company claims Phai-A1 responds 8–10 times faster than Qwen 27B. Pandaily said the comparison used the same coding-agent tasks on a single RTX 3090 GPU. The result, Pandaily added, shows neither a general speed advantage nor that the roughly 70-gigabyte BF16 weights and a 1-million-token cache fit in the card’s memory.
Phancy markets the 1-million-token context window as native. Pandaily reported that training sequences ranged from 16,000 to 1 million tokens, with YaRN scaling at a factor of four enabled throughout post-training. YaRN is a method for extending a model’s context window. The documentation warns that long contexts put significant strain on memory and the cache.
Phai-M1 generates video from text or an image, together with speech and sound. The official website says a 15-second 720p video took about three minutes on a 16-gigabyte consumer GPU. The figure comes from the company’s own tests. M1 has an online demo, but Pandaily found no corresponding download listing for it.
Hugging Face metadata lists Phai-A1’s licence as Apache 2.0, but the card says the Qwen licence must also be followed. Pandaily verified only partial downloads, not the full download or whether the model works. The models can be tried at phai.phanthy.com.
Read more: Models