Laptop achieves Opus 4.6 performance? The video’s comment section isn’t buying it

Alex Finn posted a video saying he got Qwen 3.8 running on a laptop, with performance equivalent to Opus 4.6—local, unlimited, and private.

The device is a Microsoft Surface Ultra, using an NVIDIA RTX Spark chip. In replies, Finn explained that Microsoft gave him the machine, but the company didn’t pay him or ask for content; he made the video because he liked using it.

In the video, he demonstrates how to use Hermes to load local models on whatever device you have on hand, and what kinds of projects you can build. As for the “Opus 4.6 performance” claim, someone in the comments directly asked: on which benchmark? Another user put it more bluntly: I doubt it’s even close to Opus 4.6—in fact, it’s far off.

That disagreement is more interesting than the conclusion itself. Progress in small local models is real, but claims like “equal to some cloud flagship” are hard to sell without a specific evaluation standard.

In the same thread, user ash provided more concrete data. He ran Bonsai 2 27B on an M4 MacBook Air with 24GB of memory at about 35 tokens/s, more than 4× faster than before. Bonsai 2 27B is based on Qwen3.8 27B; according to PrismML’s release notes, it is one-ninth the size of the full-precision version while retaining 98.2% on overall benchmarks.

Not everyone had a smooth experience. One reply said it didn’t run well at all on their M4 iMac with 32GB of memory. There are still plenty of variables between device specs and real-world experience.

In his own replies, ash mentioned that he currently uses Bonsai 2 27B as a sub-agent, with the orchestrator still at 6.1, and that he wants to try more local CUA scenarios next.

Finn also apologized in the replies, saying the audio didn’t turn out well this time. He hadn’t used his usual equipment and thought a $300 DJI microphone would be enough, but it wasn’t. He added that he’d have to talk to his local model to see if it could recommend a better mic. When someone asked whether Qwen is Chinese, he replied: I think Freddie Mercury is British.

Back to local models. Their value isn’t in the spec sheet; it’s in making unlimited calls and keeping data on-device available options. As for whether those options can now replace cloud flagships, the comment section clearly hasn’t reached a consensus. One user put it in a more measured way: Opus 4.6 itself is a game-changer, and if something at that level can run locally, what we’ll be able to run 12 months from now is worth imagining.