On: DeepSeek-R1

[whereismyacc]

The weights are quantized down to fewer bits in order to save on memory. The quantization loss is going to result in worse generations.

[ColonelPhantom → whereismyacc]

Ollama serves multiple versions, you can get Q8_0 from it too:

ollama run deepseek-r1:8b-llama-distill-q8_0

The real value from the unsloth ones is that they were uploaded before R1 appeared on Ollama's model list.

[AS04 → ColonelPhantom]

Unsloth also works very diligently to find and fix tokenizer issues and many other problems as soon as they can. I have comparatively little trust on ollama following up and updating everything in a timely manner. Last I checked, there is little information on when the GGUFs and etc. on ollama were updated or what llama.cpp version / git commit did they use for it. As such, quality can vary and be significantly lower with the ollama versions for new models I believe.