Please add support for Ministral-3 models

[vevi33]

Hi,
Firstly thank you very much for keeping koboldccp awesome and updated!
The new Mistral 3 models are out, but after my testing they don't work on the latest koboldccp build.
I've tested this:
https://huggingface.co/unsloth/Ministral-3-14B-Reasoning-2512-GGUF

Thanks in advance!

[vevi33]

I tried it with AMD RX 7800XT with VULKAN. I don't know if it's affecting nvidia GPU-s.

[h0ru5]

support is added in upstream llama.cpp commit cd3c118

[henk717]

For KoboldCpp other than chat adapters on our side *which shouldn't be necessary for mistral as they usually don't change format) and of course the usual testing we do when new models come out we inherent support whenever we sync up to llamacpp which is done every release. So unless a model is really quirky and unpopular to the point we don't notice you generally don't have to ask us to add a model, we do that by default even if not asked.

Considering the requests for the model though i'd like to point to our experimental builds, we usually hold KoboldCpp builds off until all the work is done and we consider it stable enough for the general public while experimental builds will have (potentially broken) support right after KoboldCpp is synced up.

Keep an eye on : concedo_experimental...ggml-org:llama.cpp:master which lists everything from llamacpp we do not yet have in our experimental build (You can ignore everything that says server: KoboldCpp does not use llamacpp's server code).

Once you see that mistral is no longer in that list you can begin to try test builds published here (assuming they were published afterwards): https://github.com/LostRuins/koboldcpp/actions

This will let you try and give feedback on early KoboldCpp development builds before we consider it stable enough, but as always its predicted to land in the next KoboldCpp release.

[MerijnHendriks]

Thank you very much!
I'm looking forward to the release, can't wait to try out the Ministral 3B model on my laptop.
If I'm not mistaken, releases are usually a month apart.
Would make for a great christmas gift!

[henk717]

Its actually a when its done approach. Depends on what happens upstream and some outside factors. Last release was delayed by weeks as lostruins laptop broke and he had to wait for it to be fixed before he could properly test the PR's people submitted.

When we know theres models queued up people want we generally try to focus on upstream support merges and a shorter release cycle. While during model draught we can go longer in between focusing more on features.

So whenever its stable enough to release again it will be released formally. But the action builds will habe it sooner.

[LostRuins]

Hello, try v1.103 which is now released!

[MerijnHendriks]

Instruct models will work fine for basic tasks, but tool calling and reasoning requires additional support to function properly:
ggml-org#17713

[henk717]

What you posted isn't relevant for koboldcpp.
Our own tool calling works in a fundamentally different way and is model agnostic.