Tsinghua and Microsoft researchers cut DeepSeek Harness token use by 88% with HarnessMM

Tsinghua and Microsoft researchers introduced HarnessMM, a method that cuts token usage by 88% when running DeepSeek models through the DeepSeek Harness framework, according to Crypto Briefing. The efficiency gain could substantially lower computational costs for large-scale AI inference, a key expense driver for AI labs and cloud providers. Reduced token consumption also improves scalability, allowing more workloads on existing GPU infrastructure. The development matters for AI-linked crypto tokens and decentralized compute networks, which often price in demand for inference capacity. Watch for adoption of HarnessMM by major inference providers and any follow-up benchmarks on cost savings per million tokens.