Strata runs ~120B Qwen 3.8 Flash Next on a 2060 laptop at 10 tok/s

On October 3, 2026, a post appeared on Reddit's LocalLLaMA forum describing a laptop with a 2060 GPU, 32 GB of RAM and 6 GB of VRAM running Qwen 3.8 Flash Next, a model of roughly 120B parameters, quantized to q2_0, on the Strata engine. According to the author, decoding came in at about 10 tok/s at a 50k context. The setup used a 4 bit KV cache and no vision.

The author said he had expected prompt processing of around 30 token/s and decode of at most 3 tok/s, given that the full Qwen 3.8 Next has around 120B parameters (not counting Engrams) and the machine has just 32 GB of RAM. He had assumed the run would crawl to a halt because of SSD swapping. The measured 10 token/s at 50k context struck him as remarkable for such an old device paired with such a large model, and he called the engine the real deal. So far the matter rests with this single post; no further developments have been reported.