cathuriges

/

tigerlily-r3

like 2

Follow

Mine of Cathuriges 7

Wanted to just mention that this includes the MMPROJ Vision Project(s), Q8_0 is already uploaded and my slow connection is working on finishing the upload for the F16 and F32 versions as well.

I appreciate it, but models here are not intended for people to actually use 😭
This is one reagent for an ongoing project, I'll make a post when it's actually finished
(You don't have to delete anything, public stuff is public lol, just letting you know x3)

Ah undertstood! 😆 I'll leave it up as you stated this would be okay to do,
but I'll wait to convert any other versions of this until you announce a full release,

Thanks btw! It's nice to "meet" you.

Hey @inflatebot sorry to bother you (and please tell me if I am)!

BUT: It's kinda crazy and awesome that there's been 551 downloads
for the GGUF Quants release I made within the first day!

(Obviously Mradermacher, being a much more renowned and respected member,
has more downloads of this model for their GGUF conversion and quantization -
BUT that's only in regards to their iMatrix release version.
The standardized Quant downloads seem to be decently popular with the release I made.
Personally? I thought that was pretty cool and worth poking ya.)

Okay, I'll stop tagging you and bothering you. Thanks again 😄
(Sorry for bugging you, honestly - I just get excited...
It's something I've been working to curb - as to not bother people)

Yeah I guess it is doing some numbers, huh? Honestly, I'm glad people seem to like it. It gives me hope that the final project (which I've put aside due to health issues but would like to return to) might be well-received. It's not quite where I want it yet, and morale is pretty low, but if all that comes of it is Tigerlily and people are into that, that's cool.

Hey there Inflatebot,

It's certainly doing some numbers as you put it! I'm really impressed actually, especially with the consistent interest and hits for the GGUF/Quantized release. I'm very sorry to hear about your health concerns, this is very regrettable as your work and model releases are very beneficial to the AI community - I mean, let's face it... Mag-Mell-R1 is probaby one of the most popular LLMs, and it's from 2024 yet still in use often coming into 2026.

Probably later today/tomorrow I'll requantize my conversion and release of Tigerlily-R3 as I've found some better quantization patterns and settings to offer a higher quality Quant without sacrificing the size of the typical quant variant.

I have been working on seeing about adding Mistral-Small-3.1's vision tower from the safetensors themselves and carefully adding them into Mag-Mell-R1. Not replacing weights obviously but rather adding so it would technically be a little larger than a 12B with the vision tower included. Then if I can update the attention weights, keys, and shapes correctly - you'd have a vision capable version of that. But, we'll see if my crazy aspirations pan out. I've already been able to do similar reworks of models but nothing quite that extensive in regards to a modified Mistral-Nemo that was merged with other models being updated with vision from technically a non-similar model base.

P.S. Is there any way I can help out with your project?

I'm very sorry to hear about your health concerns, this is very regrettable as your work and model releases are very beneficial to the AI community,

I did them around 'em! Just that sometimes it's harder to handle than others. They're both mental and physical in nature, and triggers of the former tend to make the latter worse and it's a whole thing. I'll be alright!

I mean, let's face it... Mag-Mell-R1 is probaby one of the most popular LLMs, and it's from 2024 yet still in use often coming into 2026.

Nah, people are running circles around me these days. And they deserve it! I'm not in it for popularity, I'm just in it to fill gaps where I see them. This is all a hobby for me anyway. :3

I have been working on seeing about adding Mistral-Small-3.1's vision tower from the safetensors themselves and carefully adding them into Mag-Mell-R1.

I don't have any intuition for how this would work, and I would assume it simply doesn't.
There's methodology for attaching vision towers to pretrained models after the fact, see LLaVA, but they all involve in-depth continued training to produce an adapter that works.

Not to say you shouldn't try! I'd be nothing if I didn't follow through on ideas that seemed crazy. And if it does, that'd be something haha.

P.S. Is there any way I can help out with your project?

Make or encourage good Gemma3 models. The architecture has had something of a re-evaluation with the advent of MPOA, but it's likely to still need deep tuning to account for the filtered pretrain. Lots of folks out there who could push that forward but don't seem interested, likely because of the relative unpopularity of Gemma3 compared to Nemo in the space. But it's a capable architecture and I wanna see it reach its full potential :3