Rock Matrix open-sources EdgeRock-4B, a 4B model with fixed-size reasoning memory

1. 8:01 AM
   🚨 AI News | TestingCatalog@testingcatalogX

   Rock Matrix open-sources EdgeRock-4B, a 4B model with fixed-size memory

   AIRock Matrix has open-sourced EdgeRock-4B, a small text-and-image reasoning model that keeps its attention KV cache fixed in size. The model holds the last 1,024 tokens in full and stores important older details in a separate bank, so memory use does not grow with long chats or long reasoning. Rock Matrix reports 82.9% mean pass@1 on AIME 2024–2026, compared with 18.8% for the strongest bounded-state model it tested.

   

1. 10:52 PM
   PandailyNews

   Rock Matrix open-sources EdgeRock-4B, cutting Qwen3.5-4B's 256K KV cache to 72 MiB

   AIRock Matrix, with HKUST and Carnegie Mellon University, has open-sourced EdgeRock-4B, a model converted from Alibaba's Qwen3.5-4B. Its KV cache is fixed at 72 MiB at 256K context, against 8 GiB for Qwen3.5-4B, a 99.1% cut. The team reports 6.42 times Qwen's throughput at its batch limit and 96.3% of Qwen's AIME score, with weights released under a license it says allows commercial use.

2. 2:45 AM
   QbitAINews

   EdgeRock-4B open-sources fixed-memory reasoning with 99.1% smaller KV Cache

   AIRockmatrix AI, under Yanshan Technology, and collaborators including HKUST and CMU open-sourced EdgeRock-4B, which keeps KV Cache at 72 MiB even at 256K context. The article says a single 80GB GPU runs 512 concurrent long-reasoning sequences at 12,124 tokens/s, 6.42 times Qwen3.5-4B's best, while averaging 82.9% on AIME 2024–2026, according to the report.

1. 1:55 PM
   Santiago@svpinoX

   EdgeRock-4B keeps fixed-size memory during long reasoning

   AIEdgeRock-4B is an open-weight 4B model that handles text and images and keeps its reasoning history at a fixed size. It keeps recent information, a compressed summary, and learned selections of what to retain, so memory use stays predictable as conversations grow. The post notes trade-offs: the model may drop important information, accuracy varies by task, and deployment needs compatible software.

   

The trend compares only continuously observed participants, so its scope may be smaller than the current heat total. Hover or tap for hourly heat; use arrow keys with a keyboard.