Skip to main content
Arka stores draft settings without claiming that every runtime supports them. The active vLLM, Exo, Ollama, or MLX build must expose speculative-decoding options when launched. Verify generated output quality and latency after enabling it.