A 4090 locally deployed "AI Magic Mirror": Who is the most handsome man within one meter?
I deployed an "AI magic mirror" on a 4090 - ask her "Who is the most handsome man within a radius of one meter", and she will really look at you before answering. The entire magic mirror is implemented based on my open source project VoxEMW. The model division of labor is as follows: • Eyes: MiniCPM-V 4.6 (1.3B end-to-side multimodal) • Mouth: VoxCPM2 (2B…
This AI stack turns deepseek into a full agent
Vision, tool use, and 1M context let it see, reason, and build.
China built a AI that nukes the entire OCR industry overnight.
They open-sourced a peanut-sized OCR that parses entire 100-page PDFs in one shot, with zero API calls and zero cloud costs. It's called Unlimited-OCR, a vision-language model built to handle long-horizon document
SGLANG 23: Expert Parallelism - IV
Expert weight routing. Asymmetry in choosing experts can cause high utilization in some GPUs and thus can create a bottleneck. SGL solves this by rebalancing the expert allocation across the cluster every few steps (set using
You have reached the end of the archive
All of deepseek