Dispatch
Together AI's open-source stack breakdown names five swappable layers and the specific models worth pinning to each
Published 2026-09-09, the piece decomposes an open-model workflow into model, inference, gateway/router, harness, and tools, arguing that keeping the layers independent lets you swap a new open model in minutes rather than rebuilding. It names concrete picks: Kimi K3 (1.8T total, 104B active) for the large slot and GLM 5.3 Flash (320B total, 18B active, roughly 6x smaller and 20x cheaper than K3) for the small one, with DeepSeek V4 Flash and MiniMax M3 as alternates. Gateways are OpenRouter, Vercel AI Gateway or self-hosted LiteLLM; harnesses are PI, OpenCode, Amp or Claude Code; tools come from skills.sh and mcp.so. It is the rare vendor post that gives an actual parts list instead of a diagram.
↳ Follow the thread