Flexan (user information)
Who is Flexan?
I'm Flexan. I'm a Software Developer student that likes to work with AI (and especially LLMs) in my free time. I'm especially interested in original ideas projected onto LLMs. I adore people who think of new ideas to make LLMs more useful in specific areas.
What is up with all the -GGUF models?
Since shortly I've begun to convert Hugging Face safetensors models to GGUF to allow other people (and myself) to use them early.
I specifically target Text Generation models within a range of 0B to 12B parameters that do not have a GGUF equivalent provided yet. This gives other users access to try out these models in software like vLLM, LM Studio, llamafile, etc. early without having to convert the models themselves or wait for other users to do it.
I do not convert every model, though. Any finetune of the large base models (Qwen, Gemma, Mistral) that does not bring substantial original value will not get converted by me. I convert models that are original or experimental and new.
How do you convert them?
I made a little script that converts them into the quantizations I want and uploads them with a generated README. I usually only quantize to Q2_K, Q4_K_M and Q8_0 first, but if the model is <4B, all other quants (and integer quants) will usually become available later.
The rest is not automated. I find the models to quantize by myself. I do this by literally going to the Hugging Face models page, selecting the filters, and sorting on newest models uploaded. There are quite a few cool models popping up once in a while :]
How do I request a model to be converted?
Go to the community tab of this repo, or any repo that has -GGUF at the end of it and start a discussion.
Please do not request models in the Blake-XTM repos, just to keep it organized.
Keep in mind I am not awake 24/7 and 7 days a week. I usually only convert models on Monday to Friday in the morning/afternoon.
What is Blake-XTM (Arc)?
Blake-XTM is the name I give my own models. Blake-XTM Arc is a series of models trained to call tools and optionally think (depending on which model version you pick). I initially trained it because there was a lack of reasoning models that didn't think the whole context away for a simple question (I'm looking at you, DeepSeek). I'll admit, the models failed miserably. But, that was to be expected. They were my first models I've trained, the base models were non-reasoning, and one of the models was even trained on a non-instruct base model, which is quite destructive for the small dataset I had.
Will there be more Blake-XTM models?
Yes, I am planning to release more models myself. The most likely ideas that may get released are (sorted by priority):
Blake-XTM Chatter: A silly LLM, mimicking the style of me as an online user.- An easily programmable and adaptable LLM
- A model that calculates math using inline tags.