Subscribe to download
Deepseek R1 Temp-Fix

It's under constant attack now since it gained popularity

I don't know how you managed to make R1 work. January 28-29, Openrouter, throws errors. Probably because of the high load.

On close inspection, it has an even stronger habit of doing whatever it wants and describing additional actions unprompted than other models though. I'll just try to say something to someone, which it will describe, but then it'll suddenly decide I want to run over and attack some random NPC in the area or something. I have an action suffix that explicitly tells it not to do this as well

It's really nice. It includes a level of detail I've never seen with other models. Has a habit of giving me empty responses for story text sometimes though (might be refusing to respond intentionally, as I got a complaint about graphic content one time, idk). Its a big shame we can't see the reasoning when using it in the game though. And I wish there was a slightly cheaper version I could use that doesn't suffer from the problems that the llama distil introduces. Llama sucks for story writing imo, despite being really good at RP the few times I tried.

Guess I'll have to see how a smaller version I can run locally compares if/when the dev adds support for detecting thinking tokens.

Very cool. I'll see how it goes another day when I have some time, then.

Going to give 0.6 a try though, because even though I'm fluent in English, I have to think a lot harder to picture what's going on due to there being too much detail-- a good problem to have

It's a very high reading level, but I like it. Turning the temp down may make it easier to read, I'm on 0.68 temp which has insane levels of vivid, imaginative detail. It's WAY more creative, and honestly is book-quality.

You can see its reasoning take shape for example on stuff like innate abilities. On one of my worlds it's a fallout 4 clone where you wake up in a post apocalyptic bunker, cryogenic pod, all that stuff. The AI gave me a few cryo/ice spells as a result, which was pretty cool, but I didn't want that kind of story, so I added in the player background that my body wasn't transformed in any way. Then the abilites were standard combat talents but nicely written, like the tier 1 "Dodge sideways - Roll sharply to evade projectiles or lunging attacks". I can only imagine that the reasoning model would prevent me from dodging some types of attacks using that ability unless I rolled high.

Having it do multiple tasks in one prompt is an interesting idea, didn't think of that. Though it might be difficult for the game to properly pick out the answers without a load of convoluted and specific instructions to format them in certain ways. I dunno.

Anyway, maybe I'll give it a proper try, though I don't want to spend too much money. Does it seem more or less creative than the standard assistant models? Can it 'reason out' things like being told to be more creative/less predictable?

Also, something really magical that happens with this model is that the NPCs are wayyy more involved in the story. They talk and are involved in the scene without you prompting it.

Yeah openrouter API calls don't show you the reasoning so it works fine with AIR. You just have to give it time to think since the game sends several prompts to it whenever you do anything.
Perhaps with a prompt overhaul you could just tick a box to let the smart reasoning AI handle all of it in one prompt, massively cutting down on it