What Hardware Do You Need to Run Full DeepSeek Locally? Community Weighs In
whoami-233 · reddit · 2026-08-01
A developer asked the community for the most affordable and maintainable hardware solution to run the full-weight DeepSeek model locally.
Key requirements include: generation speeds of at least 25-30 tokens/s per request, supporting at least 8 concurrent requests (ideally 16), and handling agentic tasks with long contexts of around 200k tokens. The author is currently considering a dual DGX Spark setup and is seeking better hardware alternatives from the community.
More from Infra
- PyTorch 2.13 Brings FlexAttention to Apple Silicon, Cuts Peak Memory by 4× — PyTorch · 2026-08-01
- Meta Engineer Shares MLSys Keynote: Using AI to Liberate Systems Researchers — salykova_ · 2026-08-01
- Amazon Stock Soars Record 15%, Prediction Markets Bet 2026 Capex Over $200B — Polymarket · 2026-08-01
- StudyFetch Cuts AI Inference Costs ~10x with NVIDIA Riva — nvidia · 2026-08-01
- Building a 4x3090 Local AI Workstation: Still Worth It in 2026? — davyjones10Y · 2026-08-01
- Intel Opens Up CDNA5 Instruction Set Architecture — xiaosun86 · 2026-08-01