Rock Matrix open-sources EdgeRock-4B, cutting 256K KV cache from 8 GiB to 72 MiB

Rock Matrix, with HKUST and Carnegie Mellon, has open-sourced EdgeRock-4B, built from Alibaba's Qwen3.5-4B by replacing its 8 full-attention layers with sliding-window attention over 1,024 tokens plus Rock Memory Anchors. The KV cache at 256K context drops from 8 GiB to 72 MiB (99.1%), and a single 80 GB Hopper GPU serves 512 streams at 12,124 tokens/s versus 1,889 for the base model.

• KV cache fixed at 72 MiB versus 8 GiB for Qwen3.5-4B, a 99.1% cut
• 512 concurrent streams on one 80 GB Hopper at 12,124 tokens/s, 6.42x faster
• Retains 96.3% of Qwen's AIME score and 97.5% on MATH-500
• Weights released under the EdgeRock license, which allows commercial use