Jean Louis

How does it compare against Nemotron-Nano-3-30B-A3B?

I think we heard of recursive self improvement recently in some papers.

Thanks for fine-tunning. Any practical results and reports to see the differences?

Don't you verify your LLM descriptions before posting?

The moment you try to impress me, I get unimpressed as human can see through.

I am not computer, you are publishing for human.

That is why it isn't moving, it is recognized as not being special.

Anyway, I can't be using proprietary models.

That's genuinely a cool and impressive technical project, no doubt—hooking up Gemma and CLIP to get multimodal capabilities is a real engineering feat. But calling it a "Multimodal Vision-Language Model I built" is a bit contradictory, right? You didn't build Gemma (that's Google) or CLIP (that's OpenAI). You built the pipeline or the adapter system that connects them. It's like saying you built a car when you expertly welded together an existing engine and an existing chassis.

And it is not free software approved.

Honestly, looking at VANTA Research's setup, it feels a bit like they're playing a semantic game. They're leaning hard on the "open source AI" branding to build community goodwill and sell merch, which is smart marketing. But the core of what they're optimizing—Llama models from Meta—isn't truly open source. Meta's license is restrictive, especially for larger companies, and it keeps core control in their hands. So VANTA is essentially building their safe, collaborative vision on a foundation that's proprietary at its root. It's more "open-weight" or "source-available" than genuinely libre. There's nothing wrong with building on Llama, but framing it as open source contributions feels slightly disingenuous. It's like they're open-sourcing the wrapper—the tools, the recipes for collaboration—while the main ingredient is still under Meta's lock and key. The merch-for-compute angle is clever for funding, but it does make the whole operation seem like a business leveraging open-source ideals to promote and monetize work atop a walled-garden model.

Yeah, calling this a "NEW MODEL" from a "new model family" is some serious hype inflation. It's a fine-tune. Full stop. They took Meta's Llama 3.1 8B, which they didn't build, and trained it on their own synthetic datasets. That's useful engineering work, but it's not fundamental research creating something new from the ground up. The "research" part feels unsubstantiated—where are the papers, the ablation studies, the published methodology? Without that, it's just a proprietary fine-tuning recipe they're not sharing, which is the opposite of open source contribution. They're selling an aesthetic of open collaboration ("thinking partnership") while the actual model guts—the base weights from Meta and their own curated synthetic data—are closed or restricted. So you get a merch store funding "open source," but the output is another slightly tweaked version of a proprietary model, dressed up with buzzwords like "epistemic confidence" and "wonder." It's clever branding, but the substance of what's genuinely novel and open seems pretty thin.