https://huggingface.co/benjaminradio/K2-Think-V2-GGUF
https://huggingface.co/benjaminradio/K2-Think-V2-GGUF
This one is an oddball, as you've probably guessed. From my very brief testing, I like it very lots due to smarts; it is a custom (based on llama architecture) somewhat-fresh reasoning model that is 73B of monolithic parameters. And even disregarding my (and not only mine; seen a bit of praise) personal impression, it seem rather interesting (~70B non-MoE that is not LLama 3.x nor Qwen 2.5). To my very humble understanding, the thing uses tokenizer that's a bit custom (and support for which, alas, seem not in plans for llama.cpp; tho feel free to correct me if I'm wrong), but it still possible to bolt-on a tokenizer from a similar model, which the GGUF in the link has done. But the GGUF in the link is not imatrix nor has anything beyond Q6_K and F32 formats. Even for my own testing I've converted Q6_K to IQ4_XS to fit the model into my limited hardware.
And so, if that is alright (I think would be very nice :), please ~ can you download the F32 GGUF (or if you're into extra work, the original safetensors + doing tokenizer trick that the linked GGUF mini-set describes) and turn it into the full array of GGUFs (imatrix would be ideal)? Thank you for your time!