Model types and files¶
Discover capability types¶
The JSON response names each model type and includes an explicit executable field. A category's presence in the catalog does not mean its runtime is implemented.
| Model type | Current execution path | Required files |
|---|---|---|
| Text generation | run, chat, serve |
Compatible GGUF |
| Image understanding | run --image, supported HTTP image input |
Compatible vision-language GGUF and projector |
| MIDI continuation | music |
Compatible Aria GGUF and MIDI input |
| Speech recognition | Additional backend required | Tested speech model package |
| Speech synthesis | Additional backend required | Tested voice/model package |
| Image generation | Additional backend required | Tested image-generation package |
| Video generation | Additional backend required | Tested video-generation package |
Speech recognition and synthesis are the next integration priority. They are not executable commands in CLI 0.2.2. Downloading an arbitrary checkpoint does not add a missing runtime backend.
Language-model compatibility¶
The engine has support for compatible members of the Llama, Qwen, Mistral, TinyLlama, Gemma and MedGemma families. Support is specific to architecture, quantization, backend and release. A new upstream model with a familiar family name may still require adaptation.
Use this to inspect metadata before running. Keep the exact model file and its licence information with your experiment record.
Download sources¶
The Spum Hugging Face repository currently distributes the application and CLI, not model weights. This documentation does not claim that third-party or private adapted weights have been published there.
Download model weights from their publisher or a properly licensed conversion. Follow access agreements and preserve the model card. Do not enter account tokens into a public command example. A Hugging Face repository can require separate permission even when the runtime itself is available.
Model publishing policy¶
A downloadable model entry will be added only after its artifact is released and tested with the corresponding Spum version. Each released entry should state model type, filename, source, licence, size, checksum, required companion files and a tested command. Restricted training data and private medical images are not distribution artifacts.
There is no spum pull or automatic model downloader in CLI 0.2.2. The standalone application has a separate model-management workflow. Use local file paths for CLI runs.