Xiaomi Claims 80% Compute Cost Cut for Unreleased MiMo V3

Xiaomi's HySparse2 design for the unreleased MiMo V3 model claims a 4.5x memory reduction and 80% lower compute costs. DeepSeek's V4.1 Flash uses layer note-sharing and 4x per-token memory compression; independent verification of long-context recall remains pending.

• HySparse2: 4.5x less memory, 80% lower compute costs
• DeepSeek V4.1 Flash cuts per-token memory 4x
• Xiaomi claims recall accuracy up to 256,000 tokens
• MiMo V3 is unreleased with no independent testing