Gemma4-26B distributed across 15 $55 Android TV boxes at 0.39 tok/s

SporksInjected · reddit · 2026-10-13

A demonstration of running Gemma4-26B-A4B distributed over 15 cheap off-brand Android TV boxes ($55 each): each box holds two of the model's 30 layers, networked together, producing 0.39 tokens per second. Absurd but real proof that commodity smart-TV hardware can serve as a distributed inference cluster for a capable model. The poster jokes about everyone "buying up smart TVs."

Original post →

More from Fun

• Anti-AI Skeptic Uses AI to Tailor Resume, Lands an Interview — iamKierraD · 2026-10-13
• Sentry CEO rants: Apple's cross-device sign-in flow adds endless steps — zeeg · 2026-10-13
• Jason Rohrer demos what happens when a laser pawn promotes — jasonrohrer · 2026-10-13
• Polymarket puts 67% odds on AI-developed skin cancer vaccine winning FDA approval — Polymarket · 2026-10-13
• Replies are turning into AI slop: screenshot shows bot-style engagement comments — martyamark · 2026-10-13
• The 'every AI engineer' skill ladder meme ends with a single step: Claude — mdancho84 · 2026-10-13

Fun channel →