Four 3090s: 160-175 tok/s single decode, three 262K sessions

KieranSep 212 stars

Showcases long-context concurrent serving on used cards.

Preview of Four 3090s: 160-175 tok/s single decode, three 262K sessions

Launch: Qwen3.8-Flash-Next
Use case: Gaming gpu
Format: Repo
Source: github.com

Kieran

Home lab + self-host enjoyer · local AI · virtualisation, networking · hardware enthusiast · building and breaking my own infrastructure.

See all

sergqwer

@sergqwer · Sep 29

Adds NVIDIA's NVFP4 quantisation to Strata so a single 12GB+ RTX 20 to 50-series card can host the MoE.

DominikBucko

@DominikBucko · Aug 30

Dual-3090 build with 128GB RAM reporting about 4,200 tok/s prefill and 111 tok/s decode on long prompts.

Niko1221

@Niko1221 · Sep 24

Open-source inference engine with an OpenAI/Anthropic-style local API that fits the 125B MoE onto a 12GB consumer GPU plus system RAM. The project behind the model's local-running surge.

MiaAI-Lab

@MiaAI-Lab · Sep 4

Step-by-step recipe to squeeze the 125B model onto one DGX Spark instead of two, with config and measured throughput.

Open on GitHub

Four 3090s: 160-175 tok/s single decode, three 262K sessions

More built with Qwen3.8-Flash-Next