A 3x Faster Local LLM That Can't Finish the Job — Two Days of MoE vs Dense on Real Tasks

The MoE coding model I put on my RTX 5090 was 3x faster with equal single-shot quality, so it became my opencode default. Then it went 0-for-5 on multi-step agent work. A record of porting RFC 6902 with a 112-case conformance suite as the machine judge, and the structural trade-off it exposed: the source of MoE speed and the cause of its failures are the same thing.