Turning off thinking over API: GLM-5.3 and Kimi K3 can't, and one Qwen endpoint silently ignores the switch
my-coffee-where · reddit · 2026-10-18
A Reddit deep-dive maps how each open reasoning model handles disabling thinking over API as of October 2026: Qwen uses enablethinking:false, DeepSeek uses thinking:{"type":"disabled"}, while GLM-5.3 now errors on disabled and Kimi K3 keeps thinking always-on — only reasoningeffort with low as the floor.
Key gotchas covered:
- A hosted Qwen3.6-35B-A3B endpoint accepted the off switch but silently kept thinking (159 output tokens vs 48 elsewhere), never producing a parseable answer within budget
- The /nothink soft switch is officially unsupported on Qwen 3.5/3.6/3.8
- Verify shutdown via empty reasoningcontent, a sharp completiontokens drop, or a maxtokens=16 probe
- Qwen non-thinking mode needs different sampling (temp 0.7, topp 0.8, presencepenalty 1.5); Qwen 3.8's preservethinking bloats multi-turn context; DeepSeek requires passing back reasoningcontent in tool calls or 400s
- Reasoning tokens bill as output, the pricier side of every rate card
More from coding & agent
- Clicky MCP Server queries web analytics via natural language — modelcontextprotocol · 2026-10-18
- CipherHUB ships 31 MCP cryptography tools spanning PQC and KMS — modelcontextprotocol · 2026-10-18
- Codex's new tab completion predicts your next move, and users love it — mark_k · 2026-10-18
- Open-source voxel editor ships MCP server letting Claude build and self-check 3D models — AbaGuy17 · 2026-10-18
- Dev ships AI-written game to 3 portals in an afternoon, Opus 5.5 did all the coding — Roger_M_Taylor · 2026-10-18
- 10 open-source repos to build a one-person AI content factory, from scripts to voiceovers — Roger_M_Taylor · 2026-10-18