Is there still strong interest in a dense 9b model?

I have a full model, it's ready to train. It's ~9b parameters. 9.4b to be more exact. That includes a 1/2/3 Engram table, Moonshot's AttnRes modeling, and RoPE / NoPE layering at 3:1 as more or less validated by most major labs. It uses the Llama 3 series tokenizer and LM Head as an initial start. The data fed in is logit level extraction from a Llama 3 teaching model. I've already run the first training steps to test that the model is stable, etc. I'm willing to sit and do the pre-IT training on the model. I don't know what task people really wanna do with this thing to be exact. So the focus of the IT training is a bit more vague other than giving it "Thinking" as per one of the open standards. Honestly? I don't care what people want to use it for, just that they want to use it. I figured I'd just give it into the ether and LocalLlama was a place I figure I could easily give it to. In theory the model should be more capable than any of the ~9b's we have running around with enough training. I'm also NOT a lab, so I don't have their training budgets. I basically ran all of the data production, etc on a 4090 + rented hardware. The training code is deeply optimized to run on an RTX 6000 Pro series card. A single card. All my engram research was being done before the Qwen model dropped. Qwen showed I only needed a single table injected at a layer, versus the 2 I used. 1 gave the majority of the benefit over 2. The code would be open source, the data, all of it. IDGAF. It's technically already open source as it's all in public repos at the moment. I just stare at it and question if it's worth the time. At the least I'll throw a couple hundred at it to build a "functional" pre-IT model and build the IT dataset. It will need more pretraining before the IT set probably. It's kind of unknown because no one publishes the exact figures on this type of training. If someone has a datacenter contact with a system they want to allow it to run on, we can all have it as a public model we watch. I tentatively named it "Budget" but... Localllama can name it whatever they want. It's been fun to write the full end to end, generate data, etc. Even found issues in vLLM and reported to their repo that might end up helping you guys anyways. There was some prompt loading code that could be ~10x to ~100x faster I gave examples of to them. If you read this far, thanks, Signed some ML dude who reads too many research papers and has too much spare time. edit; I went to review some Apache 2.0 licensing issues and noted a glaring hole in using Llama's tokenizer OR data. You can't even use synthetic data from their models without some licensing. I'll just have to rewrite the target to use OLMo 3 series I think. I guess I'll be back in a week after data generation and code fixes. Open source uncensored Goon model or what? I'm trying to find a useful niche to develop the model so it gets used and ends up more than a research artifact. I'm planning on building the research artifact, it's a matter of whether I can find a group of users that will actually want to use it. It's all free, so I'm not trying to monetize you. The code and data lives on GH/HF.