72 probes show Qwen3.8-Max ignores effort levels, OpenRouter quirks found
PawelHuryn · x · 2026-09-12
Pawel Huryn ran 72 probes to compare Qwen3.8-Max via OpenRouter vs Alibaba's official endpoint after seeing a score gap (19 vs 26):
- Hypothesis was OpenRouter not serving the reasoning effort level properly, as seen with other models.
- Finding: the effort level currently has zero impact on this model — it's inert on both endpoints.
- However, routing through OpenRouter and using budgettokens instead of effort causes the model's thinking to collapse entirely.
- This doesn't explain results differing 5 weeks apart, but is a notable finding on third-party routing compatibility.
Related event: Qwen3.8-Max Scores Drop on OpenRouter as effort Parameter Fails(2 posts)→
More from Models
- Granite Switch 4.1.1 ships: update models like you update software — krvarshney · 2026-10-31
- Mistral Open-Sources Le Chonk: Top Open Model on FrontierCode, Cyberdefense-Tested by CrowdStrike and Cloudflare — tensorqt · 2026-10-31
- Stanford CS336 Lecture Explains Why DPO Compares Against a Reference, Not Two Current Probabilities — stanfordnlp · 2026-10-31
- ik_llama.cpp PR #2808 adds native runtime for China Telecom's Xing4.0-29B-A4B MoE with mHC and MLA — Any_Bus9205 · 2026-10-31
- Abacus AI lets you 'vibe-code' a custom LLM from one prompt, claims 10-100x cheaper — bindureddy · 2026-10-31
- Rumor: OpenAI likely to release a very powerful model next week — imjustnewatai · 2026-10-31