Can an M5 Max Code Locally? Qwen3.8 Flash Next Tested
I left Qwen3.8-Flash-Next running on my M5 Max MacBook Pro for nine hours and fourteen minutes. In this video, I open the games, interactive apps and bug-fixing results it produced, including the original flight game that failed. If you're researching a MacBook for local AI coding, this is a real example of the output and the waiting time from one tested setup. Inspect the exact run, generated pages and published metrics: https://visualforgebench.com/run/?id=20261005T075233_qwen3.8-flash-next-gsq-rco_5a03 SETUP • Apple M5 Max MacBook Pro, 128 GB unified memory • LM Studio + Pi coding agent • Qwen3.8-Flash-Next-GSQ-RCO, IQ3_S GGUF • Loaded context: 262,144 tokens • Saved effort setting: default; this is not a matched high-effort comparison with cloud models RUN STATISTICS • 9 tests: 8 standard tasks + a separate creativity task • Run time: 9h 14m • Output: 887,939 tokens, including 389,871 thinking tokens • Overall run throughput: 26.7 output tokens/sec, including agent overhead; not peak decoding speed • Standard-task average: 5.9/10; five-task core leaderboard average: 4.9/10 • Creativity: 6.7/10, reported separately • Judge: Claude Opus 5.5, using the saved engineering rubric JET ODYSSEY REPAIR The original Qwen output scored 1.7/10 and failed to render. Claude Code repaired it after the run in 3 rounds, changing 51 of 2,706 lines. The repaired demonstration scored 6.5/10 and is clearly labelled in the video. Its repair score does not replace the original in the averages or ranking. Local Qwen inference has no model API charge here. Hardware, electricity and separate Claude judging/repair usage are not included in that statement. Older leaderboard entries use different test conditions, so the current ranking is not a controlled claim that this model beats a frontier model. CHAPTERS 0:00 Nine hours of local coding 0:33 The setup 1:26 What the numbers mean 2:15 Rotor Rampage: the hook 3:01 AeroLab 3:45 Jet Odyssey: what broke 4:22 Jet Odyssey: repaired demonstration 4:53 Blast Island 5:37 Launch Day 6:23 Alex Optics 7:10 Afterburner 7:52 Hangar Planner: fixing existing code 8:36 The creativity task 9:14 How to read the scores 10:10 Was the wait worth it? 10:56 Next: cloud supervisor, local worker Next, I'll test a frontier model supervising local Qwen coding in MaiaChat. I'll check reliability, time and cloud usage rather than assume the combination works. Subscribe if you'd like to see that follow-up, and tell me which task you'd like me to test next. MaiaChat: https://maiachat.com Model sources: https://huggingface.co/Qwen/Qwen3.8-Flash-Next https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF This video uses my HeyGen cloned voice over my actual screen recording. The instrumental soundtrack was generated in Gemini for this edit. The clock/subscribe graphics are original, and the thumbnail layouts are AI-assisted illustrations based on my face and the actual app results. #Qwen #LocalAI #M5Max