Unconfirmed social media posts from account @testingcatalog claim that Rock Matrix has open-sourced EdgeRock-4B, a small text-and-image reasoning model with fixed-capacity memory, on 31 October. @testingcatalog on X1
The posts describe a model that keeps its attention KV cache fixed in size by holding the last 1,024 tokens in full and saving important older details to a separate bank. @testingcatalog on X1 A second post said EdgeRock-4B is based on Qwen3.5-4B, with its eight full-attention layers converted to a fixed-capacity history design combining recurrent state, local attention, and Rock Memory Anchors (RMA) for retrieving selected distant details. @testingcatalog on X2
The same posts reported benchmark results. According to the account, Rock Matrix reported an 82.9% mean pass@1 on AIME 2024–2026, compared to 18.8% for the strongest bounded-state model it tested. @testingcatalog on X1 In a long-context test with inputs from 16K to 130K tokens, the model reached 94.1% query accuracy, compared to 0.4% when RMA was turned off. The attention KV cache stays capped at 72 MiB per conversation, while Qwen3.5-4B's grows to 8 GiB at 256K tokens, the post said. @testingcatalog on X2
A third post claimed local-run performance: on a MacBook Pro (M5 Pro, 28GB RAM) with Rock Matrix's PyTorch runtime on the Mac's GPU, it solved 17 × 23 step by step, started answering after 4.6 seconds, and generated about 9 tokens per second. @testingcatalog on X3
The posts cite no independent testing, Rock Matrix original benchmark methodology, prompt settings, tokenization rules, model revision, or exact baseline configuration. The MacBook result covers one arithmetic example, one time-to-first-token figure, and an approximate generation rate, without a comparison against other 4B-class models on the same hardware.
At publication time, the available evidence contains no direct statement from Rock Matrix and no primary documentation linked from the company itself. The companies named in the posts have said nothing in the supplied evidence.