← All models

Qwen3 30B A3B Instruct 2507 FP8

Alibaba · 30.5B parameters

Qwen3-30B-A3B-Instruct-2507-FP8 is Alibaba's separately released FP8 checkpoint of the non-thinking Qwen3-30B-A3B-Instruct-2507 model, using fine-grained E4M3 quantization with 128x128 block scales while the embeddings, normalization layers, router gates, and output head stay in BF16. Excluding its 1,824,768 block-scale elements, it holds the same 30,532,122,624 parameters as the BF16 release; Qwen reports 3.3B activated only at rounded precision. Its model card documents 262,144 tokens natively, and the later 1M-token DCA/MInference configuration was added only to the BF16 repository; it does not generate <think> blocks.

Benchmark scores