Joint evaluation of a sparse Qwen3-8B model

SparseMind and H3C are evaluating structured sparse training and hardware adaptation for Qwen3-8B.

Model quality

In 2:4 sparsity, two positions in each group of four are retained in selected weight matrices. The trained sparse model reaches held-out perplexity of 11.00, compared with 14.21 for the original Qwen3-8B reference. The evaluation covers 1,047,552 prediction targets after 6,000 updates. Lower perplexity indicates better text prediction. The change reflects adaptation, sparsification and training together.

Weight storage and deployment

For the eligible FP16/BF16 weight matrices in the report, 2:4 compression with position metadata requires 56.25% of dense storage, a calculated saving of 43.75%. Resource requirements for the full model, cache and operators must be measured separately for each batch size and workload.

Next steps

Further research targets 27B-scale models and system efficiency. Model scaling, end-to-end speed and energy gains remain subjects for future validation.