RockMatrix and University Partners Release EdgeRock-4B With Fixed-Size KV Cache

What happened

RockMatrix, together with HKUST, CMU, and other universities, released EdgeRock-4B with open model weights and inference code for text and image reasoning. Adapted from Qwen3.5-4B, its Rock Memory Anchors mechanism keeps attention KV Cache at 72 MiB, a reported 99.1% reduction at 256K context. In reported tests with 32K input and 32K output on one Hopper 80GB GPU, it delivered 12124 tokens/s at 512 concurrent sequences, 6.42 times the source model’s throughput, while retaining 96.3% of its average AIME 2024–2026 pass rate.

Read next

• More changes in today’s briefingNvidia open-sources Flora with projected robot video labeling costs up to 125x lower than GPT-6 Astra

The story so far

RockMatrix, together with HKUST, CMU, and other universities, released EdgeRock-4B for text and image reasoning, making its model weights and inference code publicly available for commercial use. Adapted from Qwen3.5-4B, the model uses a mechanism called Rock Memory Anchors to keep its attention KV cache at 72 MiB—a reported 99.1% reduction at a 256K context length. In reported tests using 32K input tokens and 32K output tokens on a single Hopper GPU with 80GB of memory, EdgeRock-4B reached 12,124 tokens per second across 512 concurrent sequences, or 6.42 times the source model’s throughput. The report also said it retained 96.3% of the source model’s average pass rate on AIME 2024–2026.