Fixing Overthinking

Overthinking is a Prompt Injection As many of you, I really don't like DS overthinking mere "hello"s. So I asked DeepSeek I dug up what the reasoning efforts did, because ALL seemed to overthink, for me. DeepSeek (Pro and Flash) append an extra effort guide to the system prompt. Details here , on the official Readmes. As you see, high rants about ABSOLUTE MAXIMUM and max goes full BEYOND MAXIMUM - no wonder the poor thing goes in circles for simple stuff. Its forced to overthink. low is the good one, it does not inject anything and so DeepSeek thinks as much as it needs. These injections are enforced by the chat template/encoder, which is something that can be edited serverside. But this means the official DS api DOES enforce these injections and you should be aware of it. I bet others like Ollama-Cloud, Opencode Go etc keep the official encoder. Harnesses may rob you of low You get no low in certain harnesses, so you're stuck with ABSOLUTE MAXIMUM, and I'm fairly sure most of you don't need that. I'll bring the proof: pi supports low for flash but not Pro . Opencode2 (beta) relies on models.dev , which has no low for Pro either. Other harnesses like hermes (and DS's own harness) are fine at a glance. So this is not provider-related, thank god. You can just fix the harness. Others like GLM and Minimax have their own shenanigans, I encourage you to investigate if you use them, but their effort configs aren't as drastic as DS.