I ran DeepSeek R1 671B Model against Anthropic Claude 3.5 Sonnet LLM for my dark fantasy.

• R1: Grittier tone.
• Claude: Better structure. DeepSeek needs strict negative constraints to kill the flowery fluff.

Coming back to this an hour later... ^ This. Also, I have to respectfully disagree with the previous posters about focusing so much on those negative constraints or just fiddling with the top-p settings. In my experience over the last decade of writing with different tools, the more you tell an LLM dont do this, the more it seems to obsess over that exact thing. Its like telling a kid not to think of a blue elephant. When I was wrapping up my last grimdark series using DeepSeek V3 API, I noticed that the reasoning models like R1 actually over-analyze the subtext. That is precisely why you are getting those annoying moral summaries at the end. They are literally built to explain themselves. For that gritty Hemingway vibe, I actually suggest sticking to the standard DeepSeek V3 671B without the extra reasoning layers. It is much more reliable for staying in the now of a scene and wont drift into preaching as often. One quick tip: stop trying to code the behavior in the system block and just drop two or three paragraphs of your best gritty prose as a style guide right in the prompt. It mirrors your rhythm way better than a list of instructions ever will. Plus, DeepSeek is so cheap that you can afford to run three versions of a scene and just pick the best one. Keep your context window under 4k tokens though... if the history gets too long, the model starts to hallucinate those flowery transitions again because it loses its place.

I had a rough time tuning R1. Messed with temperature too much and it started preaching... be careful with those settings, maybe try lowering Top-P instead to keep it from getting weird.

Like someone mentioned, keeping things simple is a good baseline, but I actually have to disagree about stripping the prompts back too far! DeepSeek R1 is such an amazing beast for dark fantasy if you just feed it the right anchors. Tbh, the trick is to use style anchors instead of just negative constraints. Few-shot examples work way better than rules for this model. Basically, just paste a page of your best gritty prose and tell the model to mirror that specific rhythm. Ngl, NovelCrafter Personal Plan is a total lifesaver for this because it lets you pin style references so the model doesnt wander off into AI-narrator land. This setup honestly beats using OpenAI GPT-4o for the heavy, gritty stuff once you get the prompt dialed in. Another fantastic option is Rexy AI Writing Tool if you want that snappy dialogue without the monologues. Just keep showing it exactly what you want! It's honestly so satisfying when the prose finally hits that Hemingway vibe... you're gonna crush those last 10k words!

Like someone mentioned, these models can get a bit preachy if you push them too hard. I have been using DeepSeek for a while now and honestly, keeping it simple has been the only way I can stay sane. Whenever I tried to micromanage the Hemingway style or the deep POV, the model would eventually just break and start giving me those weird moral endings again. It feels like the more complex the system prompt gets, the less reliable the output becomes. Based on my trial and error, here are a few things I would be careful about:

• Stay away from long lists of negative constraints because it usually confuses the logic.
• Be careful with the temperature settings; anything too high makes the prose way too experimental and purple.
• Try to avoid generic instructions like be gritty and use specific examples of the tone you want instead. It is a decent model for the price but definitely requires a light touch. You dont want to kill the consistency by over-steering the ship... hope you can get those last 10k words done tho.