Running four high-level OpenCode agents + subagents with mlx_lm.server continuous batching and MiniMax M2.5 (6-bit). Fits easily on a 512GB M3 Ultra. Generation is quite fast. But prefill is still slow compared to cloud servers.