Joint evaluation of a sparse Qwen3-8B model

SparseMind and H3C are evaluating structured sparse training and hardware adaptation for Qwen3-8B. H3C provides engineering support and an eight-H20 test environment.

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

Under the report’s FP16/BF16 eligible-weight format, 2:4 compression including position metadata uses 56.25% of dense storage, a calculated saving of 43.75%. Resources for the full model, cache, operators, batches and workloads must be measured separately.

Quantization can be combined with sparsity, subject to model-quality and hardware validation. Further evaluation will measure memory use, throughput, latency, energy and deployment conditions.

Next steps

The team will continue validating reusable model and deployment configurations and investigate extension to 27B-scale models. That extension and system-level speed and energy gains remain future work.