TheDrummer

/

Snowpiercer-15B-v4-GGUF

like 12

skill issue

default settings work fine. or you can read and learn.

https://rpwithai.com/understanding-sampler-settings-for-ai-roleplay/

The only issue with this version of Snowpiecer is the reduced context length from 65536 to 2048 although I generally use 8192 as my default. As for performance and quality its thedrummer, all of the models are superb !!

The only issue with this version of Snowpiecer is the reduced context length from 65536 to 2048 although I generally use 8192 as my default. As for performance and quality its thedrummer, all of the models are superb !!

Where do you see that? Based on the config.json of the model, it's set to 65k tokens. https://huggingface.co/TheDrummer/Snowpiercer-15B-v4/blob/main/config.json

I'm not super familiar with smaller mistral models so i can't say when things start to devolve with them, but i would assume 8-16k should be fine like their bigger 24B variants.

I'd also assume the model uses Mistral-V7-Tekken for Context/Instruct templates. That's what i generally use for the 24B's and it works well for me.

the context is not reduced to 2048. must be something on your end.

https://huggingface.co/TheDrummer/Snowpiercer-15B-v4/blob/main/config.json

it uses ChatML

For inference I am using LM Studio (0.3.34 todays updated version with Vulkan 1.62) and the max context length for this specific model version is 2048. snowpiercer-15b-v3c.gguf, Q8-0. I am not a geek or expert when it comes to Ai and I do not know of anyway that I can increase the context length beyond the default values available.

For inference I am using LM Studio (0.3.34 todays updated version with Vulkan 1.62) and the max context length for this specific model version is 2048. snowpiercer-15b-v3c.gguf, Q8-0. I am not a geek or expert when it comes to Ai and I do not know of anyway that I can increase the context length beyond the default values available.

You may want to try KoboldCpp. That's what i use, and it is pretty customizable and very easy to use. https://github.com/LostRuins/koboldcpp . It ties into Sillytavern just fine.

When you build your first config that works for you, make sure you also specify your admin dir where you keep your RP gguf models and save your config in that directory. That will allow you to go to the "admin" button in the KoboldCpp web ui and swap between other models and config files. I generally keep a few different default configs with kv sizes and context sizes so i can swap them as i move between larger and smaller models. Please note that it will not see subdirectories so you must have all your configs and gguf files together.

come visit us in the beaver discord. we are friendly and happy to help :)

You are rude, get help.

can confirm, beaver discord is fun

Sometimes you need to look at your own issues first. The model is excellent. If you’re getting something for free and still complaining, that’s unnecessary.

I have it deployed on HF and trying to increase the default Max token value from 8k to 32k but it keeps reverting at 8k only. Is there a way to force it to 32k token size?

I have it deployed on HF and trying to increase the default Max token value from 8k to 32k but it keeps reverting at 8k only. Is there a way to force it to 32k token size?

that looks like a huggingface problem. Load it in your system locally and set whatever context size you want.

I have to comment: this has been my absolute favorite LLM for story writing by a long shot and for a crazy long time now. I keep trying new models hoping to see snazzier word choice, more robustness in settings, perfect logic, but Snowpiercer never fails to disappoint. Whereas it delivers with no fuss, others run far slower for comparable quality, want to :think> at random intervals without serious hand holding, reject mundane topics, start spewing gibberish far more often, and just always dissapoint me more. It has a few areas where it could be improved, as its logic is not perfect and it might write something like 'they talked to the teller at the bank on the 4th of July' missing that it'd be closed, but it constantly suprises me how much it gets right. I also second using this with Koboldcpp as that is by far my favorite interface for its ease of use. I use Temp=1, Top-P=0.95,Rep Pen 1.02, TopK=0 and Smoothing Factor 1.5 and Context 131k and it never gives gibberish and doen't fall into repetition for pages and pages. Other settings run great too. If someone has an easier model with higher lexical diversity that doesn't fight the user, I'd love to see it suggested. I've tried over 200 models of various sizes and formats and this one makes me the happiest. Thanks to all who made it possible and special thanks to The Drummer: you are a hero and I have no idea what the OP is on.

I was being too clever and organised and it turns out LM Studio does not like you using certain characters when naming and creating folders. I have learned a lot since December 2025 and just finished upgrading my AI workstation with 32Gb VRam so it is now a very happy bunny. You are right Gutts, it takes a lot and something special to beat the impish range of models from Sica Rius.