DeepSeek open-sourced the production V4 Flash on July 31. One week later, developer Ryan Zhou published an open-source deployment stack that serves the full model — 284B-parameter MoE, 13B active per token, original FP8 weights, no quantization — on a single AMD MI300X. It hit 344 points on Hacker News, the day’s top AI post.
What the repo actually is
Not a model, a recipe: Docker Compose stack, pinned file overlays, patches against upstream inference code, and tuning tables. The MI300X’s 192GB of HBM swallows the weights whole; a ~235GB CPU RAM tier handles KV cache for the 1M-token context. Zhou had to fix FP8 handling, MoE routing under concurrency, and several untuned ROCm kernels to make it stable.
Why this matters
The story isn’t the chip, it’s the floor. A frontier-class open model that needed a cluster last week now fits on one card — an AMD card, sidestepping the NVIDIA supply queue entirely. For teams that want DeepSeek private — regulated industries, on-prem enterprise — the ask just shrank from a cluster to a single GPU. Chinese open weights on non-NVIDIA hardware is now a production path, not a demo.
You Might Also Like
- Title ds4 Deepseek v4 Metal Local Inference Engine by Antirez Redis Creator Runs v4 Flash on a Single Macbook
- Deepseek v4 pro v4 Flash Ship With 1m Context and 0 28 m Output
- Deepseek v4 Flash 0731 Hits 82 7 on Terminal Bench Chasing Opus 4 8 at 0 14 m Tokens
- Lunair Turns a Single Sentence Into a Fully Animated Explainer Video and it Actually Works
- Ricursive Intelligence Just Raised 335m in 4 Months and i Think Nvidia Should be Paying Attention
Top AI Product 
Leave a comment