Something that's not appreciated nearly enough about Deepseek: the FIM Completion endpoint. You can easily take advantage of it on Windows using AutoHotkey.

If you don't know, DeepSeek is currently the only AI lab that's offering a FIM Completion endpoint that's not mainly for coding. What's special about this is that instead of talking to the model through the usual user/assistant chat format, you basically just give it some text and it just keeps it going, or give it text before + after a certain point and it generates what should go in the middle. This seems to activate the model's pure next-text prediction ability, which is wildly fun for creative writing, storytelling, worldbuilding, editing, and all sorts of stuff. I've been playing around with this a lot recently while working on AhkLLM , a free and open-source Windows LLM app I've been building in AutoHotkey v2. AhkLLM has two commands for this, FIM Fill and FIM Continue. Fill gets the text before + after your cursor, Continue only gets the text before it, and since AhkLLM uses UI Automation to grab and inject text directly in a lot of Windows apps you can use it pretty much wherever you're already writing, without having to copy the text into a chat, explain what you want, copy the answer back etc. So you can do something completely stupid like: My wife found the hidden folder on my desktop. [DEEPSEEK WILL WRITE THE MISSING MIDDLE HERE] Anyway, I own a strip club in Malta now. and see what kind of batshit middle it'll come up with to connect the two. What's also fun is keeping the beginning exactly the same and only changing the ending, then running it again. You can make the ending increasingly stupid and Deepseek will still try to somehow make the middle work, which I've wasted a ton of fun time on. For actual writing though, this is undoubtedly way more useful than a typical chat workflow when you're already mid-paragraph and just want to see where the model thinks things could go next, or quickly edit a few words in the middle of a sentence and have the model rewrite it while keeping the surrounding text intact. The writing itself is lot better than the same model via its chat interface, since the model doesn't have to pretend to be a "helpful" assistant and can just directly continue what you're writing. I think completions kind of got forgotten once everything moved to chat, which is a real shame because base completion models are really good writers when you can just give them text and let them cook. I remember people were doing crazy stuff with the original GPT-3 completions back in the day, and Deepseek brings that back to a large extent, and it's dirt cheap in comparison. Anyway, both FIM Fill and FIM Continue are built into AhkLLM , and you can assign them to whatever hotkeys you want under Settings -> Commands. AhkLLM itself is a much bigger app at this point with a full chat GUI and a bunch of other stuff, but I mainly wanted to post this because I almost never see anyone talking about Deepseek's FIM endpoint for normal text. I'm curious if anyone else has played around with FIM or text continuation in general for creative writing, and if so, what did you think of it? If you're a Windows user and want to give it a quick try, you can download Ahkllm here: github.com/AmitTzah/ahkllm/releases/latest/download/AhkLLM.zip