Flash-Next at 65 tok/s on a 12GB RTX 5070 with a purpose-built engine

Went from 15 tok/s in llama.cpp to about 65 tok/s by writing an inference engine for this one model and this class of PC, with 2-bit RCO-GSQ quants running faster still.

Open on Reddit

Launch: Qwen3.8-Flash-Next
Use case: Gaming gpu
Format: Build
Source: reddit.com

KnownAd4832

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 Reddit

Flash-Next at 65 tok/s on a 12GB RTX 5070 with a purpose-built engine

More built with Qwen3.8-Flash-Next