Qwen3.8 Flash Next hits 60tok/s on RTX 3060 12GB with Strata

The writer's test sits against a background he describes as rising graphics card prices, partly driven by local AI model use. His point is that models once confined to server-class GPUs now run on consumer hardware, which he frames as a shift in how people might use AI.

Within the same article, he also ran a second model, Qwen3.8 27B GSQ-RCO, on the identical PC via llama.cpp, reaching around 20tok/s. That comparison puts the Strata result in context: the engine, not just the model, is doing much of the work.

He adds a caveat from several days of use: local models are not yet a replacement for paid services. He suggests using them with a specific purpose in mind, such as handling confidential information or avoiding cloud outages, rather than as a general substitute.