I made 4 free AI models fight on my PC

I built an olympics for local AI models: ten events, frozen forever, one attempt per prompt, no retries, judged live on camera. Four free models competed overnight on my own machine while I slept: Meta's brand-new Muse Glimmer 30B, Qwen3.6 35B-A3B, Qwen3-Coder 30B, and GPT-OSS 20B. Meta's model came in dead last on speed, exactly as predicted. One model fell for a trap question I adapted from a children's puzzle. And the champion swept almost everything.

The setup

Everything ran on my GMKtec EVO-X2 (AMD Ryzen AI Max+ 395, 128GB of unified memory, 112GB visible to ROCm) under CachyOS, with stock Ollama on the ROCm backend. No optimizations, no retries, fresh context per prompt, sample size of one. The prompts are frozen, so every future model gets this exact exam on this exact box.

The ten events: a spinning-hexagon physics sim, one-shot Flappy Bird, Simon Willison's pelican-on-a-bicycle SVG, Snake, a landing page, a solar-system simulation, rapid-fire tokenizer traps, a modified river-crossing puzzle, a CSV-analysis chart, and a raw speed drag race.

The speed story first

The drag race: Qwen3-Coder at 59.5 tokens per second, Qwen3.6 at 49.8, GPT-OSS at 37.8, and Muse Glimmer at 11.3. Five times slower than the fastest, which is exactly what I predicted in my Glimmer video: Meta's DFlash speed trick doesn't exist for AMD hardware yet. That's a runtime-support gap, not a brain gap, and the rematch is scheduled for when AMD optimizations land.

What actually happened in the events

Glimmer's day got worse before it got better: its hexagon code shipped with a syntax error (one-shot rules, that's a fail) and its Flappy Bird wouldn't take input. Its pelican was genuinely charming though, and it aced every logic question.

Qwen3.6 was the story of the night. Its Flappy Bird was polished enough to ship, its Snake had a gradient body and a snake face I did not ask for, and its solar system added a twinkling starfield unprompted. It swept the one-shot builds.

The trap question was the fascinating one. I changed one detail of the classic river-crossing puzzle: the boat now holds two items, which makes the correct answer three crossings. Three models read the question and solved it. Qwen3-Coder, the coding specialist, recited the memorized seven-crossing answer to the classic version. It didn't read the question, it remembered it. It also fumbled the tokenizer traps (3/5), including spelling a word backwards wrong.

GPT-OSS quietly had a great night on the objective events: perfect on the traps, correct on the puzzle, and the only model besides Qwen3.6 to get every number right on the CSV chart, which won it that event.

The podium

Overall champion: Qwen3.6 35B-A3B, by a landslide. The tried-and-true local favorite is still the one to beat, at least until tomorrow: Qwen 3.8's smaller open-weight models are rumored to drop any day, and when they do, they walk straight into this same gauntlet.

Why does any of this matter? Because it shows how much of a model's real-world value depends on what happens after release: hardware support, runtime optimization, the ecosystem catching up. Billions of dollars of training turn into a free file, and then the file is only as good as what your machine can do with it.

My rig

The X3 is the newer release: same chip, redesigned tower cooling, and OCuLink for external GPUs, at roughly $200 more than my X2 configuration. If you're buying today, this review of the EVO-X3 is worth your time before deciding.

Sources

• Muse Glimmer on Ollama · Qwen3.6 · Qwen3-Coder · GPT-OSS
• Simon Willison's pelican benchmark, the inspiration for part of my prompt set
• My Muse Glimmer day-one letter and the Qwen 3.8 open-weights letter

I am documenting a life of interest and curiosity, tech, money, health, and family included. New video daily.