A user has released a custom GGUF quantization of the Qwen 3.5 122B model, named "Heretic," utilizing ROCmFP4 and iMatrix techniques.

  • The checkpoint is designed for Strix Halo hardware with 10 billion active parameters.
  • It requires approximately 60.70 GiB of memory.
  • Benchmarks indicate a performance of 28.45 tokens per second in BF16 precision.
  • The model achieves a perplexity (KLD) of 0.100716 and a pipeline throughput of 353.3 tokens per second.