Warning: Qwen3.8 settings prone to typos, thinking mode differs from v3.6
misanthrophiccunt · reddit · 2026-08-20
A user warns the community that Qwen3.8 settings are easily confused with version 3.6. Specifically, the "thinking for precise coding tasks" settings from the Unsloth guide apply to the thinking mode in 3.6, but similar-looking parameters in 3.8 are actually for the non-thinking (Instruct) mode. Applying the wrong settings prevents the model from working as intended.
More from Models
- OpenMaMMUT Collection: Weights on DataComp and Re-LAION Available — wightmanr · 2026-08-20
- Open-Weight Model Bets on Recursive Self-Critique for Improvement — omarsar0 · 2026-08-20
- dots3-note Demonstrates Interaction, Hypothesis Testing, and Memory Updates — omarsar0 · 2026-08-20
- dots3-note Preview Shows Adaptation in Unseen Environments — omarsar0 · 2026-08-20
- dots3-note Preview: 280B Open-Weight Model with Self-Evaluation for Long Tasks — omarsar0 · 2026-08-20
- Bloomberg pits seven leading AI agents against each other in a vibe-coding challenge — pstAsiatech · 2026-08-20