Local DeepSeek Setup Tested on Three GPUs
Jaswanth04 · reddit · 2026-07-19
The author ran the DeepSeek-V4-Flash Q8 quantized version locally on an older platform: the 3 GPUs used were two RTX 3090s (24GB) and one RTX 5090 (32GB), paired with 128GB of DDR4 RAM and a Ryzen 3950X.
They provided the complete llama.cpp launch parameters, including multi-GPU layer split, tensor split, moving MOE layers to the CPU, disabling mmap, and using a 262144 context. They claimed this override tensor placement method leads to more balanced GPU utilization.
In terms of actual measured speed, prompt eval was around 37.9 tok/s, and generation was about 9.07 tok/s. The author believes it outperforms Qwen 3.5 v2 on coding tasks and asks how token speed can be further improved without upgrading hardware.
More from Infra
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27
- Gary Marcus: With Q3 data added, Anthropic's compute-efficiency growth shows deceleration — GaryMarcus · 2026-10-27
- SpaceX locks in $40B GPU financing deal with Apollo, the largest ever committed — TechCrunch AI · 2026-10-27
- Inference is just Kubernetes with a multi-trillion-dollar market, and the communities overlap — matt_slotnick · 2026-10-27
- vLLM eyes Rust migration as inference engines race to rewrite the stack — AccBalanced · 2026-10-27