Qwen has released the FP8-quantized weights for its Qwen3.5-397B-A17B model, a 397-billion-parameter causal language model with vision capabilities that activates only 17 billion parameters per inference step.
The model utilizes a hybrid architecture combining Gated Delta Networks and sparse Mixture-of-Experts to deliver high-throughput inference with minimal latency. It supports native context lengths of 262,144 tokens, extensible up to 1,010,000, and covers 201 languages and dialects.
The FP8 quantization employs a fine-grained method with a block size of 128, maintaining performance metrics nearly identical to the original model while reducing resource requirements.