DeepSeek V4.1-Flash CED + CSA2 + Engram + Single-Pass mHC + DSpark - Looped decoder (recurrent depth) + Mixture-of-Recursions routing - 15B total, ~2-3.5B active, 1M context staged See ARCHITECTURE.md for full spec. Target: frontier reasoning per active param on SWE-bench / Terminal-Bench / DeepSWE.
Independent publisher
Gugu8
Gugu8
Architectures, Training Data, Deepseek.
Models in Library2
Datasets in Library0
Models on Hugging Face6
Followers—
Models
A vision encoder with a runtime-selectable token budget, built to sit between an image and an LLM. A coarse 32px grid covers the whole image; a router spends extra 16px detail only on the regions that need it. Output length is 49 + 3k tokens — 49 at k=0, 121 at k=24 — so you choose the tokens/fidelity trade-off per image, at inference, with no retraining. 24.94M parameters. Trained in 6000 steps (~30 min) on 2× Kaggle T4 by distillation from google/siglip-base-patch16-224 (86M). This is a working, honest baseline, not a competitive encoder. On the Flickr30k official test split it reaches 8.95 mean R@1 against the teacher's 83.25 — roughly 10% of its teacher. Chance on 1000 images is 0.10.…