Benchmarking DeepSeek Locally on Dual RTX 3080s

bkin777 · reddit · 2026-07-16

A user shared benchmark results for running the quantized version of DeepSeek V4 Flash on a rig with 2× RTX 3080 20GB + 64GB DDR5, providing full hardware specs and launch parameters.

Key takeaways:

The author notes it runs stably but can still be optimized. The post also mentions that this specific fork fixes output anomalies caused by context quantization.

Related event: Developers Benchmark DeepSeek Quantized Models Locally(2 posts)→

Original post →

More from Infra

Infra channel →