Together AI: tune the system before the weights — cheapest levers first when evals fail

togethercompute · x · 2026-11-21

Together AI shares a practical playbook: when an open model misses your eval, start with the cheapest levers in order — rewrite the system prompt, tune inference settings (temperature, reasoning, sampling), upgrade context and harness (tools, skills, MCP, context), and only then change model weights via fine-tuning or distillation. The takeaway: test the system around the model before reaching for fine-tuning.

Original post →

More from coding & agent

coding & agent channel →