EdgeRock goes open source: constant-memory reasoning model with code, weights and paper
alifcoder · x · 2026-10-27
The author shared all EdgeRock resources; the project is open source on GitHub (Apache-2.0) with 672 stars.
- GitHub: rockmatrix-ai/EdgeRock includes inference and evaluation code, with vLLM and SGLang runtime support
- Hugging Face: model weights available at rockmatrix-ai/EdgeRock
- Paper & whitepaper: available on Arxiv and via the official whitepaper
Positioned as "strong reasoning at constant memory": a multimodal model with O(N) generation, O(1) persistent history, and KV cache fixed at 72 MiB even at 256K context. README ships in both English and Chinese.
Related event: EdgeRock Open-Sourced: Constant 72 MiB KV Cache for 256K Context(4 posts)→
More from Models
- Anthropic's usage restrictions spark backlash over paid-model rights and open-weight futures — QuixiAI · 2026-10-31
- Reddit user declares Claude the winner: better models and now more usage than OpenAI — IthrowUgo · 2026-10-31
- OpenAI's $100 Pro looks like a steal vs Anthropic's $200 Ultra: doubled quotas, no 5-hour window — gethackteam · 2026-10-31
- ChatGPT desktop voice mode still slow and unreliable, hands-on test finds — koltregaskes · 2026-10-31
- Google fixes Antigravity IDE to add Argon; early user calls it a good model — zacharynado · 2026-10-31
- Mistral open-sources Le Chonk, tops FrontierCode among open models — _AndrewZhao · 2026-10-31