Change default model repo to Hugging Face

[olliewalsh]

When not explicitly provided, the default model repository should change from Ollama to Hugging Face.

Ollama has diverged from upstream llama.cpp. As a result of this, an increasing number of newer ollama models are not compatible with lama.cpp (e.g gpt-oss).

While we can workaround this be using shortnames to override specific models, this is not sustainable.

[rhatdan]

Making this change requires the RamaLama project to greatly improve our handling of shortnames.conf,

We need to go through the most popular AI models, and update the shortnames.conf file to pick a default.

$ ramalama run hf://deepseek
Downloading hf:///deepseek ...
Trying to pull hf:///deepseek ...
URL pull failed and hf not available
Error: Failed to pull model: 'No metadata found

The average user getting this error is going to be at a loss and will switch very quickly to ollama.

When a new popular model hits the press, RamaLama needs to quickly release a new version.

If huggingface or llama.cpp or a different entity provided an easy API for ramalama to call into for shortnames expansion then I am 100% for this change.

[olliewalsh]

Yes, as I commented on #2405 this isn't a straightforward case of changing the default which is why I raised an issue for it.

Making this change requires the RamaLama project to greatly improve our handling of shortnames.conf,
We need to go through the most popular AI models, and update the shortnames.conf file to pick a default.

Yes, preserving the current ollama-style shortname behaviour is an issue.

When a new popular model hits the press, RamaLama needs to quickly release a new version.

Flipping that around - when a new popular model hits the press, the ollama model will not be compatible with llama.cpp (only legacy model are IIUC), so will fail and the average user will be at a loss and switch very quickly to ollama...

[akashWhoCodes]

Hi @olliewalsh @rhatdan,

Thanks for opening this issue and for your guidance on why Ollama was the original default. I agree that the divergence between Ollama and upstream llama.cpp makes the current default increasingly brittle.

I’ve opened PR #2405 which implements the switch to Hugging Face. To avoid surprising users who rely on Ollama shortnames, my plan is to stage the change across releases:

• In the next minor release we keep ollama as the default, emit a deprecation warning at start‑up and update the docs to explain how to set RAMALAMA_TRANSPORT=ollama or use an explicit ollama:// prefix to retain current behaviour.

• Meanwhile we can document or maintain short‑name mappings for popular Hugging Face models so that common shortnames remain functional.

• In a later major release we flip the default to Hugging Face once people have had time to adjust.

Please let me know if this phased approach makes sense or if you have other suggestions.