The llama.cpp library released a new version on Oct 20, 2026. It changes the mixture‑of‑experts (MoE) router so the selection no longer depends on random backend noise. The fix gives each expert its own score range, reducing output differences from 4e‑3 to under 5e‑7. The same update adds the actual Apple device name (e.g., M2 Ultra) to test tables, so CPU and GPU results can be told apart.
Why it matters
Developers using llama.cpp on Apple Silicon will see more consistent model outputs and clearer performance reports.
