How the HELL do you go so fast?
I made this AS A JOKE, and you GGUF'd it already.
The question is more how these quants already have 223 downloads in a single day. I think it's an interesting concept to subtract Qwen2-0.5B-Instruct from Qwen2.5-0.5B-Instruct but as expected the resulting model is unfortunately quite broken: "I believe the meaning of life is痣加DAC ')'ved набvol interprepriERIC番文章名字aryl族自治州齐全 pCTXiap茅百分.EventSystemsérc价值两端reo弢CDFearesimaliscilèveCTXведения濠耻gg análftware递岁以下cí…”"
It seems like mradermacher liked the concept as well and so decided to do static quants for it. Static quants are relatively cheap so the bar to get static quants of your model is relatively low and the title alone sounding interesting will likely be enough to catch mradermacher's attention in the few seconds he has to decide for each newly released model if and what quants to provide.
Regarding your question how things are going so fast I recommend you take a look at http://hf.tst.eu/status.html. It is all automated and as you can see mradermacher currently has 7 servers dedicated to compute quants. Some of them are from mradermacher or his company while others are provided by the community like nico1 from me and rich1 from RichardErkhov. Thanks to this absolute massive amount of compute power and network bandwith, we can quant models at an insane rate. While there is a queue of currently 2734 models new releases will skip the queue if mraderemacher decides he likes them or quants are explicitly requested by users.
Oh. Yes, I am aware of the model being broken, it was just an experiment.
That nico1 said, and also, sometimes models turn out to be broken, but a lot of authors don't even try their model first, or wait for a gguf quant themselves., I'll delete this model, now that we know. Thanks for notigying us, even if that wasn't you main intention :)
And yes, I am debating if I should wait a few days before quanting a model.
And yes, I am debating if I should wait a few days before quanting a model.
Most interest of a new exciting model seems to happen in the first week after its release so I don't think waiting is good option as that would leave hundreds of users having to do their own quants to try a newly released model. As you can see even this broken model already has 227 downloads. By offering quants of this model, we likely saved hundreds of hours everyone would have otherwise spent creating their own quants to try it out. I would argue even for broken models having quants might sometimes be worth it in order to not waste everyone’s time until the model is confirmed to be broken. Ideally the original author would by then either delete or put a disclaimer on the original models so users know it is broken and they don’t have to waste their time quantizing it themselves.