Phanthy Releases Phai-A1, a 1M Context Model Trained on Huawei Ascend

·2026.11.20 11:30

Key point

Phai-A1, an open-source 1M context model trained on Huawei Ascend NPUs, has been released.

1 / 4

Details

  • Phanthy released Phai-A1-35B-1M, an open-weight model post-trained on Huawei Ascend NPUs based on Qwen3.6-35B-A3B
  • Achieved 83.6% on SWE-Bench Verified, a 10.2%p improvement over the base model, and 65.0 points on SWE-Bench Pro, a 15.5%p improvement
  • Supports 1M context via YaRN and is distributed under the Apache 2.0 license
  • Provides a demo of faster generation speed compared to Dense models on a single RTX 3090 due to its MoE structure (3B active)

Technical Features and Limitations

  • MTP (Multi-Token Prediction) weights have been removed, making the base's speculative decoding settings unusable
  • When using 1M context, the KV cache occupies approximately 20GiB, making it impossible to run alongside a Q4_K_M quantized model in a 24GB VRAM environment
  • Vision encoder tensors are present, but the official GGUF lacks mmproj, so it can only be used for text-only tasks

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.