Alibaba Qwen Team Just Released FP8 Builds of Qwen3-Next-80B-A3B (Instruct & Thinking), Bringing 80B/3B-Active Hybrid-MoE to Commodity GPUs

Alibabas Qwen team has introduced FP8-quantized checkpoints for their Qwen3-Next-80B-A3B models, available in Instruct and Thinking variants, optimized for high-throughput inference with ultra-long context windows and efficient Mixture-of-Experts (MoE) architecture. These FP8 models mirror the earlier BF16 releases but incorporate fine-grained FP8 weights with a block size of 128, providing performance benefits without altering evaluation benchmarks, and include deployment notes compatible with sglang and vLLM nightly builds. The Qwen3-Next-80B-A