[Feature]: DeepSeek-R1-Distill-Qwen or similar distilled DeepSeek gguf support

[zsogitbe]

Martin, do we have support for DeepSeek-R1-Distill-Qwen gguf or similar already (distilled version of deepseek)?
Thank you.

[martindevans]

No I don't think so, not yet. We're currently about 5 weeks out of date.

Glancing through the PRs, I can see these two which looks relevant:

• llama : add support for Deepseek-R1-Qwen distill model ggerganov/llama.cpp#11310 (8 days ago)
• Add support for DeepSeek V3 ggerganov/llama.cpp#11049 (26 days ago)

The next version update is already in the works, I think there's just one major issue left to resolve which I hope to get to this weekend.

[zsogitbe]

OK! The last cpp version should add support. Thank you!

[phil-scott-78]

Pretty excited to see the new binaries added. Not just deepseek, but it'll be an opportunity to try and incorporate the changes they've made to support jinja templates and as of a few days ago the tooling support.

[martindevans]

Is the new and improved Jinja support in the core llama.cpp, or just in the examples (e.g. common.cpp)? A lot of big changes like that land in examples (which LLamaSharp does not include) before they make their way into the main project.

[phil-scott-78]

it's in core. A pretty big one too, I think some of the features related to Deepthink might require it. I know the tooling stuff does ggerganov/llama.cpp#11016