JetBrains Chose Qwen 3.6 Over 3.8 Because 3.8 Needs Reasoning, and Then Cut Generation Time in Half With MTP Plus N-Gram Speculative Decoding

JetBrains Junie blog·medium signalXBlueskyLinkedInCopy link

The companion engineering post explains that Qwen 3.8 degrades badly with reasoning off, and enabling it generates about 5x more tokens for a roughly 4x slowdown, which killed it for local M5 use. JetBrains disabled reasoning entirely on 3.6 for 2-3x fewer tokens and about 2x speed, used 4-bit over 8-bit for another 2x, applied 8-bit arithmetic during prefill on M5 for about 40% faster prefill, and combined Multi-Token Prediction with n-gram speculative decoding for a further 2x on generation. They also extended rolling context to reuse KV-cache across tasks so the agent stops re-reading files, and disabled multi-agent mode outright.