Four-day blind test: GLM-5.3-Flash beats DeepSeek V4.1 Flash for local assistant
A 7-round self-built test on two GB10 boxes via a Hermes agent: GLM won both Sonic-style games and kept the assistant job, while DeepSeek's engine crashed 5 times and took 90 minutes vs 18 on golf, though it clearly won the short-film round.
Ran a 4 day tryout for the brain behind my local AI assistant. GLM-5.3-Flash vs DeepSeek V4.1 Flash, with Ornith-1.5 joining late. Instead of picking off benchmark tables I wrote my own tests and ran them through my Hermes agent on my two GB10 boxes: games, a short film, phone golf and security posts. A bot did the measuring with real key presses and real taps, and I did the judging, blind for the film, the golf and the posts. It went 7 rounds. GLM took both Sonic-style games, and DeepSeek's AI art version crashed on a normal launch. Then DeepSeek won the short film round and it wasn't close, so it took over my assistant for a day. Then came golf. Both games passed every check, but DeepSeek's engine crashed twice and it took an hour and a half to GLM's 18 minutes. Over the whole test DeepSeek's engine stopped 5 times and GLM's never did, so GLM kept the job. It's one run each, so a close round could flip on a rerun. Every round is written up on my site and you can play the games. The golf ones work on your phone, the Sonic ones want a keyboard: https://gardenofnull.cc/portfolio/reports/brain-head-to-head/