[Feature]Unsloth Studio: Pick context lenght and models folder

[fcoberrios]

Let us users select 2 things, before loading a model, after clicking one let us select the variables of the model and its context lenght, because when you open Unsloth Studio the default settings allways load first, and after you load a model you might realize that you didnt change the context lenght so you need to offload the model, change the context lenght and load the model again.

Also, let us choose folders/files to load models we might have laying around, so that way we dont need to download models from the list and that way we can load specific models we might want to load.

Thank you for this awesome Open Source Project!!! :)

[shimmyshimmer]

Both are now available in Unsloth Studio! Please update and thanks for trying it out and your feedback!! :D

Updates: https://unsloth.ai/docs/new/changelog

[melroy89]

after clicking one let us select the variables of the model and its context lenght, because when you open Unsloth Studio the default settings allways load first, and after you load a model you might realize that you didnt change the context lenght so you need to offload the model, change the context lenght and load the model again.

I think this is still an issue to this day, right? I can not select/save the context length before loading in the model? Even trying to create a new preset for it.

On my machine this is causing a lot of issues since by default Qwen 3.6 27B 4Q gets loaded with 95488 context length with f16 cache type, using Unsloth Studio (I share my GPU also with my OS). So it almost freezes my whole system every time I load-in the model.

I want to select a lower context length and/or select KV Cache Dtype during the first time loading of the model. And make it part of the present..??

EDIT: I run version 2026.5.1

[ambhaiji]

Context length now adjustable. Keep in mind this is not needed as llama.cpp smartly uses the exact context you need via --fit on

this poorly assumes that unsloth(studio) is used on a separate system or VM with fixed limits and no other processes used its resources, meanwhile many users are like @melroy89 and I who are using in on our own personal systems and so --fit on just hijacks every bit of resource possible and kills any possibility doing anything else in parallel.

I would kindly ask the devs to look into these 3 things if possible:

1. if --fit on can tell us how much context is possible for us to get and to show us that before hand so we can set our limits based on that
2. set context limits and other model parameters before loading.
3. the option to remove the model from memory after it is loaded onto the GPU.