Unsloth Studio
Unsloth Studio runs AI models on your own graphics card and serves them the way OpenAI does. Connect it to Kitlig and those models chat with you, plan edits, write Lottie animations, translate dialogue, and make images, video, speech and music, with no per-call cost. Everything stays on your machine.
This page gets you set up quickly, then lists the models we tested and what each one is good for.
Set it up
- Install Unsloth Studio, start it, and download at least one chat model (the list below says which).
- In Kitlig, open Settings, then API Keys, then the Custom tab.
- Pick the Custom preset and click Add endpoint.
- Replace the address with your Unsloth Studio API address. It ends in
/v1, for examplehttps://your-server.example.com/v1. - If your Studio uses an API key (they start with
sk-unsloth-), paste it into API key and click Save key. - Click Test. Kitlig lists the server's models and picks the best tested one as the Default model.
- Set Tools to Essentials, tick Vision if your default model can see images, and leave Thinking on for now.
- If the server runs on another computer, click Edit next to Server hardware and type in its graphics card, VRAM, CPU and RAM.
- Open the Model tests panel and click Use best picks. The Assistant, the built-in agents, the Video Assistant and Describe & Tag now use your own models.
Note
Unsloth Studio holds one big model in memory at a time. Switching between a chat model, an image model and a video model means a load each time, which can take a minute or more. The first request after a switch is always the slowest.
Which models to use
We tested these on an RTX 3090 with 24 GB of VRAM, a Ryzen 5 5500 and 32 GB of RAM. Speeds on your machine will differ; run Test in the Model tests panel for your own numbers, and see everyone's results on the benchmark page.
Chat, the Assistant and planning
| Model | Quant | Speed | Good for |
|---|---|---|---|
| Qwen3.6-35B-A3B-MTP | UD-IQ4_NL | 136 tok/s | The best all-rounder: passed every test, including looking at pictures. Our pick for the Assistant. |
| gemma-4-31B-it (qat) | UD-Q4_K_XL | 69 tok/s | Passed every test, pictures included. Slower, but it writes long JSON (Lottie, translations) faster because it thinks less first. |
| gemma-4-26B-A4B-it (qat) | UD-Q4_K_XL | 162 tok/s | Very fast and passed every text test. It can't see pictures. |
| gemma-4-E2B-it (qat) | UD-Q4_K_XL | about 300 tok/s | Tiny and very fast, sees pictures. Good for translation and tagging; too small to trust with Assistant edits. |
| Ternary-Bonsai-27B | Q2_g64 | 64 tok/s | Passed everything, pictures included. |
A model name like 35B-A3B means a large model that only works a small part of itself per word. Those run nearly as fast as a small model while knowing much more, and at a 4-bit quant they fit in 24 GB. Dense models of the same size that don't fit in VRAM spill into system memory and slow to a crawl: several 27B models ran at 6 to 12 tok/s on our test machine.
Warning
Skip these for the Assistant: gemma-4-12B-coder writes tool calls as plain text instead of running them, and Qwen3.8-27B-OBLITERATED can't call tools at all. MiniMax-H3 is too big for a 24 GB card.
If a thinking model is slow or runs out of room in long answers, untick Thinking on the endpoint. Qwen3.8 finished a Lottie in 54 seconds with Thinking off and didn't finish at all with it on.
Images
| Model | What it does here |
|---|---|
| flux-3-image-klein-GGUF | Text to Image and Edit Image. Edits from an instruction in about 50 seconds and needs nothing extra. The best place to start. |
| MOSS-Image-SFT-GGUF | Text to Image. About 80 seconds for the first image, faster after. |
| Qwen-Image-Edit-2511-GGUF | Edit Image, but only after Unsloth Studio downloads about 27 GB of extra parts. Load it once in Studio before using it from Kitlig. |
Video
| Model | What it does here |
|---|---|
| Wan2.2-TI2V-5B-GGUF | Text to Video, or animates a Start image. A 1 second clip takes about 2 minutes on a 24 GB card, a 5 second clip about 7. Keep clips short while you try ideas. |
Speech
| Model | What it does here |
|---|---|
| Chatterbox-Turbo-GGUF | Text to Speech with its own voices. About 18 seconds per line. |
| MioTTS-1.7B-GGUF, DotTTS-Edit-GGUF | Text to Speech that can clone a voice from a Voice to clone recording. |
| Seed-VC, RVC | Speech to Speech: re-voice a recording while keeping its timing. Seed-VC copies a sample voice; RVC uses a trained voice name. |
Transcription
| Model | What it does here |
|---|---|
| Parakeet-TDT-0.6B-v3-GGUF | Transcribe with proper timed subtitles. About 2 seconds for a short clip. Our pick. |
| MOSS-Transcribe-Diarize-GGUF | Transcribe with speaker names on each line, plus a speaker list. Use it for scenes with several people talking. |
| whisper-large-v3-turbo | Transcribe, but Unsloth's Whisper returns no timings, so the subtitle file has a single line for the whole clip. |
On Automatic, Transcribe picks a model with timed subtitles first.
Music and stems
Music & Sound Effects and Stem Separation work with Unsloth's music and separation models, such as Stable Audio 3 Small Music, Stable Audio 3 Small SFX and BS-RoFormer.
Note
We haven't tested these two yet, because our test server had no music or separation model. Download one in Unsloth Studio and they show up in the Model list.
Use them in Kitlig
The media operations are in every operation picker under Your Model Server: Text to Image, Edit Image, Text to Video, Text to Speech, Speech to Speech, Transcribe, Music & Sound Effects and Stem Separation. They work in the Sandbox, Batch and the Workflow canvas, and the Assistant and connected MCP clients can run them too.
Each one has a Model list showing only the models your server has for that job. Leave it on Automatic to use a suitable model, one already loaded first. If the server turns a model down, for example because it isn't fully downloaded, Automatic moves on to the next.
When something goes wrong
Kitlig can't reach the server. Check the address ends in /v1 and that Studio is running. Click Test again.
A model is missing from a list. Download it in Unsloth Studio, then reopen the operation. Chat models appear in the Default model list after you click Test.
Edit Image stops and says the model needs more files. That model needs extra parts Unsloth Studio hasn't downloaded. Kitlig won't start a download that size on its own. Load the model once in Unsloth Studio, then try again.
Replies are slow or stop halfway. Untick Thinking on the endpoint, or pick a smaller or mixture-of-experts model from the chat table above.
See also: Settings for every Custom endpoint option, and Model Benchmarks for testing your own models.