Starterkit

[basxto]

Would it make sense to release a new starterkit now that there are many new small models like Qwen3VL, ministral3, Z-Image and Kokoro?

[LostRuins]

Sure, feel free to contribute one! I'd be happy to add it and host the .kcppt for you if it's well tested!

[basxto]

It’s not well tested, but for now I have this. It’s just the starterset with all but the whisper models swapped.

• Z Image Turbo from the template on the release page
• Kokoro model as linked on the release page
• Unsloth’s Ministral 3 3B Reasoning quant seems to work
• Official Qwen3 Embedding 0.6B seems to work

I’m not sure how to properly test the embedding model. I added some stories to the DB and ministral could quote from it when I asked about a person appearing in a story.

I prefer to set image generation (sdui and lite) to 9 steps and a low resolution, but I don’t see how that could be configured via the config. I also don’t have a feeling for how much resources it’s occupying right now, is there anything similar to ollama ps?

I loaded the Qwen3 4B Instruct of ZIT as text model, which works. They probably still get loaded into RAM twice? Though it needs less disk space. This isn’t practical right now since that’s not a VL model. It’s possible that Z Image Edit will use Qwen 3 VL, but it’s not clear if that will have a distilled version.

Features that would be missing in the starterpack:

• voice input like Omni, but whisper can convert
• image editing like Kontext
• ZIT LoRA use (doesn’t seem to download from civitai)

Tested on: Radeon RX 580 with 8GB VRAM and Ryzen 5 5600X with 32GB RAM running KoboldCPP via Vulkan on Arch

ztarterpack.kcppt.json

[LostRuins]

Thanks! I'll give it a test and add it to the collection if all goes well. Also pinging @henk717 if he wants to check it out

[LostRuins]

@basxto one potential issue: i have heard reports of the embeddings generated by Qwen3-Embedding being broken (not accurate). Have you encountered this issue? If so we might want to use a different one like BGE-m3 or something

[basxto]

I have no real comparison. I would need to do the same test with artic and see how different that is. It looked a bit randomly stitched together, but it wasn’t clear to me if that was qwen mebedding’s doing or ministral. Ministral 3b is pretty small, so it can get confused occasionally. It also regularly spoke for me when I used it via lite, so I had edit mode active often. I had no issues on lcpp, but I only asked it simple questions and no follow up questions. I forgot to test visual input

[LostRuins]

Okay, I can confirm the qwen embedding does work, but it's not size efficient. I will probably swap that with BGE

[LostRuins]

Alright, have tested and everything looks good, I've uploaded it as a new starter pack here https://huggingface.co/koboldcpp/kcppt/blob/main/starter-pack-v2.kcppt

Thanks!

[basxto]

I screwed up Ministral. As model I used Reasoning, but the mmproj of Instruct. Though I never saw it produce any thinking blocks, maybe it’s better to completely use Instruct.

[LostRuins]

i'll swap it then