DUAL STRIX HALO EXPERIMENT 24 SEP 2026 EDT
64 builds.
Two AMD desktops.
One live build storm.
Thirty-two independent apps and games per machine, generating at the same time on local AMD Ryzen AI MAX+ 395 systems. This is the complete wall: every finished artifact, ready to open and play.
A whole wall in 5:32.
One first generation pass. Two machines. Sixty-four concurrent requests.
Two machines. One output stream.
Ryzen AI MAX+ 395
112,713 first-pass output tokens
Ryzen AI MAX+ 395
107,275 first-pass output tokens
1 332.44 seconds from launch through the first generation pass. 2 Highest five-second rolling output rate; machine peaks occurred at different times and should not be added. 3 First-pass output tokens divided by the same 332.44-second wall time. Rates exclude repair work. One of 64 first-pass outputs hit its token limit; the final gallery includes the later repair and publication fixes.
Open the wall.
Every tile is a real, interactive single-file HTML artifact. Pick one and step inside.
No matches. Try another machine, type, or search.
Local from prompt to pixels.
The challenge
Sixty-four distinct briefs launched together: games, utilities, and mini-sites. Each worker produced its own self-contained HTML, CSS, and JavaScript, with no external runtime dependencies.
The stack
Two AMD Ryzen AI MAX+ 395 Strix Halo desktops, 32 concurrent requests on each. Both ran Ciru-optimized Gemma 4 26B through vLLM. The live wall displayed generation and output as the run progressed.
What the numbers mean
The speed and 5:32 time describe only the original generation pass. Sixty-three outputs ended normally; one hit its output limit. Repairs and six small publication fixes were completed afterward and are reflected in the playable gallery, not the benchmark figures.
AMD provided one of the two Ryzen AI Halo systems used for this content. The run, measurements, apps, and showcase were produced by Ciru/Crown.