I am totally obsessed with DeepSeek R1 671B Parameters lately! For coding, I compared a minimalist prompt against a heavy Chain-of-Thought one. The minimalist style is super fast but sometimes skips logic tho. The heavy one gives perfect code but loops forever. Honestly, just using a simple Direct Specialist prompt works best for me... it keeps the reasoning tight without those endless cycles. Less is definitely more here!

Stumbled on this discussion today and wanted to jump in with some cost considerations, cuz honestly, those massive multi-paragraph system prompts are just gonna eat into your budget over time. If you are hitting the DeepSeek-R1 API 671B Full Model hard for heavy refactoring, those extra input tokens really add up. I have had a lot of luck using a Constraint-Based approach instead of a Persona-Based one. Basically, instead of telling it who to be, tell it what not to do. I use something like: Prioritize logical density. Avoid repeating state transitions in thought blocks. Output Python code following PEP8. This keeps the reasoning focused without triggering those infinite loops you mentioned. It basically tells the model to stop yapping once the logic is sound. Tbh, if you want to save some serious cash while testing these prompts, you should check out OpenRouter AI API Service. They usually have great pricing and you can swap models easily to see if you can get away with a smaller context window. Also, definitely look into LiteLLM Open Source Proxy if you are managing multiple keys; it helps track exactly where your token spend is going. I found that keeping the system instructions under 100 tokens total is the sweet spot for performance vs. cost. Less is definitely more when youre paying by the million tokens.

Ive been messing with this model non-stop lately, its honestly wild. Quick question tho, are you using the official API or something like Groq Cloud LPU Inference for your Python tasks? The latency totally changes how I structure my prompts. Tbh, I found that adding an instruction to focus on modularity prevents those loops. You might also want to check out DeepSeek-R1-Distill-Llama-70B if you need faster iterations without the 671B bloat.

I absolutely love what DeepSeek-R1 is doing compared to models like GPT-4o! The logic is just fantastic and the reasoning is on another level. But seriously, be so careful with those long, complex system prompts you might be used to from OpenAI. I've found that if you try to prime it too much like you would with Claude, the reasoning engine actually trips over itself. It is a total mistake to copy-paste prompts between brands! R1 has a totally different architecture. I've seen it go into these infinite loops just because I tried to force a specific persona or added too many instructions. My big warning is to avoid any instructions that tell the model how to think. It already knows how to do that! Just tell it what the final file should look like. If you clutter the system prompt with think step by step or be a master coder, you are just asking for trouble and wasted tokens. Keep it lean or you will break the logic flow!

No way, I literally just dealt with this yesterday. Small world.

Yep, this is the way

Regarding what #3 said about the token costs, it really hits home. I have been trying to get consistent results for my own projects and the reliability is just all over the place. Its honestly exhausting to manage.

• The core issue seems to be how the model handles the reinforcement learning signals that govern those thought tags.
• Since the reasoning is intrinsic to its training, even a well-structured system prompt feels like it is fighting the models natural tendency to over-analyze.
• I have noticed that the deeper it goes into a thought loop, the more likely it is to forget the original architectural constraints I provided. It feels like we are paying for the model to argue with itself rather than doing the actual work. I am using the DeepSeek-R1 671B API for basic scripts and even then, the lack of predictability makes it hard to trust for anything production-ready. Its like the model is too smart for its own good but lacks the focus needed for a reliable developer experience.