I've experimented with a few of the models. Each model has it's own strength, so it's up to you to find a model that has a writing style you vibe with. General rule of thumb though is to avoid reasoning models. And if you want to be cost effective, just load up on lots of RAM and get a GPU to run an MoE.

Deepseek v3 0325 is good all round but loses coherence on more complex prompts

Deepseek v3.1 is better at staying coherent than Deepseek V3 0325, but it is slightly less articulate

Kimi-K2 (both versions) are very poetic but laden with purple prose, 0905 is more coherent.

Qwen3 235B22A 2507 Is very balanced, and a good choice for a local set up, you can run it at decent speeds with a modest set up of 128GB of RAM and a 16GB graphics card.

Qwen3 80B3A is very fast, but it loses coherence and support is a bit patchy right now, but the Tongyi lab are working fast to implement the architecture in various OS frameworks

GLM and GPT-OSS are not very good at creative writing, and GPT-OSS just loses track of basic creative writing.

Baidu Ernie is a step below most of the models

Grok 2 and Cohere Command A are very hard to run and honestly a generation behind these other models, so it's not worth wasting too much time on them.

If you have the hardware though, Sao10K and TheDrummer have some of the best writing finetunes on Dense models like Mistral Large 2 and Llama 3.x 70B. Euryale and Behemoth are some of the best creative writing models. And finally, can't go wrong with Midnight Miqu, model still holds up well.