While testing Sage A1.5 35B yesterday, I received feedback from others that it loops/reasons too much (thanks again @Franzferdinan57, @0xneoptr, and @Ramon317317 for sharing). I investigated using my local Sage Q3_K-exps model, then repeated the same tests on @OpenBMB MiniCPM-V 4.7 35B IQ3_M. The finding: Yes, it reasons way too much. 3.65x more on average. I ran 11 Sage probes and nine MiniCPM probes. I tried a basic definition, a 120-word explanation, simple multiplication, and modular arithmetic. I also varied sampling settings and random seeds, tested thinking-off controls, and investigated Sage’s Verification Mode. Thinking was otherwise enabled with an unlimited budget. Results are below. Every probe finished naturally without hitting my diagnostic limits. Both models answered the arithmetic correctly and respected the word limit. I found repeated planning and checking in Sage’s simple-question traces, but no sustained verbatim loops. MiniCPM generally planned briefly, and answered. So, is Sage A1.5 reasoning too much? My findings suggest yes - in most cases. Sage used substantially more reasoning on simple prompts, but that difference disappeared on our multi-step arithmetic example. This suggests its deliberative style (strong tendency around deep research, planning, and verification) may be better suited to certain tasks than others, although broader testing is needed to establish whether the additional reasoning improves results. If anyone has a more solid theory (and armed with more expertise around this model), I'd be interested to know? [Update: I have been since informed by @latent_node that Sage needs temperature 1.0, top_p 0.95, top_k 20, presence_penalty 1.5 - will try these specific conditions and come back with more]