Downloads · 30 days
18
5% of all-time downloads
SuperDan/Affine-Q3.5ft-5DHWeG9naPinHGTZp43akB6LE3ztq7xA6uWg84synAHZqEFR
Affine-Q3.5ft-5DHWeG9naPinHGTZp43akB6LE3ztq7xA6uWg84synAHZqEFR is a image-text-to-text model from SuperDan. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. It is set up for transformers. The card lists the license as apache-2.0.
"Just a little guy trying to do his best on some benchmarks."
Downloads · 30 days
18
5% of all-time downloads
All-time downloads
346
Public
Parameters
27.8B
278 GB on disk
Likes
1
Public
Click a slice to open those files.
.safetensors55.6 GB · 100%
How the weights are stored.
BF1627.8B · 100%
From the Hugging Face model README
"Just a little guy trying to do his best on some benchmarks."
Fine-tuned variant of Qwen3.5-27B. We took an already-scary model and made it slightly more unhinged in the direction of agentic reasoning and tool use. You're welcome.
Qwen3.5-27B has a genuinely weird architecture — hybrid linear recurrent attention (GatedDeltaNet) fused with sparse self-attention. Training on it requires custom infrastructure that doesn't Just Work™ out of the box. We built the infrastructure. It works. The model is better now.
The fine-tuning applies reward-shaped optimization with a custom signal targeting structured reasoning, tool-call formatting, and multi-step task completion. The result is a model that is annoyingly good at following instructions and won't hallucinate a fake API endpoint at you (usually).
We're competing on Bittensor Subnet 120 (Affine) — a subnet where validators send your model agentic coding tasks and score it on whether it actually solves them. It's basically a gladiator arena for language models, except nobody bleeds and the prize is TAO emissions.
Our model enters the arena. We believe in it. We have trained it. We have watched it learn. We are emotionally attached in a way that is probably not healthy.
The GatedDeltaNet linear attention layers in Qwen3.5 require specialized kernel support that most training frameworks quietly don't have. Running this model without the right stack means your GPU sits at 0% pretending to work while the CPU slowly suffers. We have the right stack. The GPU is very busy and very warm.
Fine-tuned by AffineGo.
Base: Qwen/Qwen3.5-27B | Precision: BF16 | Fully merged weights
No LoRA adapters were harmed in the making of this model.
Apache 2.0. Go nuts.