🚩 Report: Spam
Model is fake & spam. For instance, compare model-00008-of-01939.safetensors to OpenGVLab/InternVL3-78B model-00008-of-00033.safetensors. You can see it in the file’s Large File Pointer Details and the raw pointer: 110bc8463ce8fa4c51b9492a83761aadf9c7d5ff227d4c7b461a61eedf3c3682. The same is true for many other files here, they match exactly another model and are not the model proposed on the HF page.
Another example, shard 442
https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct/blob/main/model-00185-of-00241.safetensors
42beb029c4f6b8b57a8642471550aedffa50c3f120ae65595d3eb37f59f43ed8
Shard 621
https://huggingface.co/Qwen/Qwen3-235B-A22B-Thinking-2507/blob/main/model-00005-of-00118.safetensors
5c40ca3e04508ca33cba860330220ba183b268f144a7e38799b80841b714f985
etc, every model piece is copied from somewhere else and not an actual full model.
At the very least if DynaMoE is something that uses all these models it is not using the proper licensing for each model taken & used here (without mention of them anywhere).
The index.json is also complete nonsense and uses tensor names that dont exist: https://huggingface.co/deca-ai/3-alpha-ultra/blob/main/model.safetensors.index.json
You can also see in the commit history where they felt the need to update the total size of the safetensors index so HF reports their nonsense correctly (afaik, correct me if wrong) https://huggingface.co/deca-ai/3-alpha-ultra/commit/61ecc89970416ee8695f00979551a6d6884783bb
It's a very good lesson on how to spot either fake spam or horrifically engineered models, if anything! :)
Let me answer this:
One: this isn’t spam. Deca 3 Alpha is an experiment, and yes, it’s scaffolded from existing models. That was intentional and mentioned upfront. We’re testing routing, reproducibility, and scaling — we didn't pretrain this
two, all reused components are properly licensed. We’ll be adding a NOTICE.md to clarify provenance, including InternVL.
three, the commits you see were because HF wasn't correctly reporting the size when we added the entire model
horrifically engineered models
True, but I wouldn't say it that way. I'd say:
under-engineered models
I think it's time we understand why we have "alpha" in the name.
We are going to update the readme to reflect this. We are going to add that notice file. We are going to release everything transparently, but till then please be patient.
Show us where you run this model and some outputs.
Don't believe me? Try for yourself: https://deca.genlabs.dev/chat But it uses only one expert because everyone is in such a rush
Don't believe me? Try for yourself: https://deca.genlabs.dev/chat But it uses only one expert because everyone is in such a rush
Giving all benefit of the doubt that it was a very badly worded release and not just lying, this endpoint also doesn't work (yet).
I still don't understand how it could possibly be a scam.
You don't understand much, we've been over that already. Now go away. Be happy I don't press what i found googling your github / email trace :)
Just don't come back. Reassess your life.
YOUR MODEL IS NOTHING! IT SERVES ZERO PURPOSE! YOU SHOULD QUANT YOURSELF. NOW!
Ngl bullying a scammer was fun. (your mom looks nice, she wouldn't be proud of you, tho)
Reading through this discussion like
meta-thought: this person is clearly farming the attention they're getting from this and we should stop responding to them, not doxx them
half of their comments and all of their commits are vibe coded, there is clearly nothing of value, its not clear literally anything was trained
I think this argument is foolish, but it's understandable that only a handful of users on HF have human-level intelligence. Furthermore, the model's approach is likely dynamic routing, which I think the original poster's point is quite reasonable and entirely correct. However, dynamic routing differs significantly from dynamic MOE, a technical point that needs clarification.
I think this argument is foolish, but it's understandable that only a handful of users on HF have human-level intelligence. Furthermore, the model's approach is likely dynamic routing, which I think the original poster's point is quite reasonable and entirely correct. However, dynamic routing differs significantly from dynamic MOE, a technical point that needs clarification.
This.
Meanwhile, conversation locked and I’m putting this on LMArena (if they approve) to test