Qwen 3.5 122B Heretic ROCmFP4 iMatrix Released for AMD GPUs

Qwen 3.5 122B Heretic ROCmFP4 iMatrix has been released, featuring 122 B total parameters with 10 B active, 60.70 GiB memory usage, and a throughput of 28.45 tokens per second. The model operates in BF16 precision and reports a KLD of 0.100716, achieving a peak performance of 353.3 tokens per second in the PP benchmark. The ROCmFP4 iMatrix variant is optimized for AMD GPUs, leveraging ROCmFP4 mixed‑precision for efficient inference. The release includes a Hugging Face model card and a GGUF checkpoint, enabling easy integration into existing inference pipelines. The model’s large context window and mixture‑of‑experts architecture make it suitable for high‑throughput, low‑latency applications on AMD hardware. The release demonstrates the growing competitiveness of open‑weight models on non‑NVIDIA platforms.