Story time: Qwen3.8-Flash-Next on my Strix Halo laptop vs Claude Opus 5.5 on the same feature

For the last few weeks most of my coding has been done locally with Qwen3.8-Flash-Next, so I gave it and Opus 5.5 the same high complexity feature to build on LlamaStash (a complex and large Rust project) and compared the results.

Setup: ASUS ROG Flow Z13 (Strix Halo, 128GB), Flash-Next at xhigh effort with Pi as the harness. Opus 5.5 ran in Claude Code at medium effort. I wanted xhigh for Opus as well, but Claude changed it to medium when I picked the latest model and I didn't notice it until the task was done. But I think medium is probabbly a fairer setting anyway.

Task: add a llamastash daemon restart command that reuses the existing start and stop code. I kept the prompts vague on purpose and gave both the same prompts.

The end result was interesting. I asked GPT 5.6, Opus 5.5 and Flash-Next to review and compare both PRs (new sessions). GPT and Flash-Next picked the Flash-Next PR (#89) and Opus picked its own (#88). I also did my own review and found the Flash-Next one better as it had better tests and handled edge cases better. I ended up merging #89, after porting the fixes that the reviews picked from #88.

Keep in mind:

• Opus was on medium effort. With xhigh it would have used way more tokens, taken a bit more time and probably would have done a better implementation.
• Flash-Next ran on an older Halogen version (0.14.0), and Halogen dropped the connection once, so the last part ran on Gufo. The current Halogen does around 1,400 t/s prefill and 46 t/s decode on my laptop at 70 W, so I think the time will drop a lot if I redo the test.
• The $7.53 is what Claude Code reported for the whole Opus session, which includes a later fix to the PR. The 0.15 kWh assumes 70 W for the whole 130 minutes.

Opus is still 2 to 10 times faster and I still use it for planning and reviews. But the actual coding now happens on my laptop, and to me it is crazy that I can run a local model that can challenge a frontier model like this.

Full post with my setup, the engine benchmarks and a second task comparison: https://deepu.tech/local-ai-qwen3.8-flash-next-best-local-llm