He squeezed Qwen 3.8 into a laptop and claims it matches Opus 4.6

Alex Finn posted a demo on X: Qwen 3.8 running on a laptop. He claims performance reaches Opus 4.6, the model runs locally, is private, and has no call-count limits. The device is an RTX Spark.

A local model means the weights run on your own device, and requests do not go through a cloud API. For developers, this affects reproducibility and controllability. For ordinary users, it affects whether data leaves their machine and whether they can still use it when offline.

Video:

The laptop is only the surface. The deployment path is more worth watching. In his tweet, Alex Finn says the video will cover how to load local models on any device using Hermes, and what kinds of projects you can do with these models.

A little cold water is needed here. The bottleneck for local models is usually not the demo clip. VRAM, quantization, cooling, battery life, and initial setup—any one of them can pull 'any device' back to reality. The specific specs of the RTX Spark, the quantized version of Qwen 3.8 used, and the Hermes installation steps determine whether this setup can be replicated. Alex Finn says 'amazing'; that is his experience. When watching the demo, pay attention to these details; it is more useful than just looking at performance slogans.

If this path holds, the value is clear: data does not leave the machine, it works even offline, and long tasks do not require watching a quota. People building agents, processing documents, or assisting with code will have one more controllable option. Ordinary users can also have a little less worry about 'uploaded means trained.' Do not rush to replace cloud models. Run it once on your own machine, check latency and output stability, and then decide.

Source: Original tweet