Made for Swiss self-employed & SMEs
Most Swiss businesses assume they need a cloud subscription to use powerful AI — but the real cost is your client data leaving the building. The DS4 engine, built by Redis creator antirez in just one week, changes that: it runs frontier AI models like DeepSeek V4 Flash and GLM 5.2 entirely on your own hardware, with zero per-token costs and everything staying on premises.
What DS4 actually is
DS4 (also called DwarfStar, or ds4.c) is not a general framework — it's a single-purpose engine antirez wrote for one model family: DeepSeek V4 Flash/PRO, V4.1 Flash, GLM 5.2/5.3, and Qwen 3.8 Flash Next. It doesn't reuse llama.cpp's code but borrows its GGUF quantization ideas, under MIT license.
The trick is an asymmetric 2/8-bit quant recipe: routed experts, which make up most of the 284B parameters, get squashed to IQ2_XXS/Q4_K since each expert only ever handles a fraction of tokens — while attention, shared experts and output layers stay at sharper Q8_0. This means you can run a 284B‑parameter model on a Mac with 128 GB RAM, or on a GPU box with 256 GB+.
Why it matters for a small Swiss business
Two things usually stop a small office from using AI on its data: cost and data protection. DS4 is free and MIT‑licensed — you buy no subscription, pay no per‑token fee. The model stays on your Mac, your NAS, or your small server. Under Swiss data protection law, that is a far easier story to tell clients than sending their files to a server abroad. For doctors, notaries, fiduciaries and anyone handling sensitive documents, that is often the deciding point.
Hardware you need
- Best: Apple Silicon Mac with 128 GB RAM (Q4 quant, ~153 GB)
- Minimum: 96 GB RAM with the lightest quant (IQ2_XXS, ~81 GB)
- GPU route: NVIDIA DGX Spark/Ada-Lovelace (CUDA) or AMD Strix Halo (ROCm) for even lighter loads
If you don't have that RAM, the model won't crash — it just won't load the full 284B. You can still run smaller subsets.
A sensible way to try it
Clone the repo, build for your backend, then pull the weights from Hugging Face. Start by feeding just one document folder and asking simple questions. Use it for a week and watch how often it saves you a search through old emails and folders. Only then decide whether to roll it out more widely. That is the strongwinds way: no hype, no big bang. One real problem, one honest test, and a decision based on what you actually saw.
