EdgeRock: Open-Source Multimodal Reasoning at Constant Memory With O(1) History
xiaohu · x · 2026-10-23
rockmatrix-ai open-sourced EdgeRock (557 GitHub stars, Apache-2.0), a multimodal model design delivering "strong reasoning at constant memory": generation scales O(N) while persistent history is compressed to O(1), so context never balloons memory mid-run.
- Inference and evaluation code released, weights on Hugging Face, plus a technical report
- Ships vLLM/SGLang integration, Dockerfile, and usage docs
- Target use case: robots, autonomous vehicles, and edge hardware where memory is fixed and shared with cameras, radar, and motor control — constant-memory inference removes the risk of an LLM suddenly consuming all available RAM
Related event: EdgeRock Open-Sources Constant-Memory Multimodal Inference for Edge Devices(4 posts)→
More from Embodied
- Figure claims 81 min of robot data per second; critic calls it low-value slop — MarcJSchmidt · 2026-10-23
- FSD 14 lite driver review: near-perfect performance, HW3 is all you need — haydendevs · 2026-10-23
- First real-world sighting of a Tesla Cybercab on public roads — MatthewChang · 2026-10-23
- ETH Zurich's ACP teaches Unitree G1 whole-body ladder climbing with 13 minutes of demos — CyberRobooo · 2026-10-23
- NVIDIA GTC robotics roundup: Agile ONE SN humanoid, Isaac Lab 3.0, and Jensen's 'ChatGPT moment' — lukas_m_ziegler · 2026-10-23
- Tesla patent reveals flat-printing and thermoforming process for mass-producing robot fingertip touch sensors — CyberRobooo · 2026-10-23