Autoregressive base model, general text. Deeper/narrower architecture than v2 with QK-Norm + XSA attention changes and a digit-atomic tokenizer. Improves coherence (BLiMP) and perplexity over v2; ARC-Easy knowledge recall is flat as a disclosed tradeoff. Not instruction-tuned — continues text, doesn't chat. Arithmetic is still frequently wrong in free-form generation.

Run the same prompt and settings through both autoregressive base models at once to see the effect of v3's deeper/narrower architecture and attention changes, plus the benchmark deltas below. (v1, Coder-v1, and the diffusion models aren't included here — try them individually in the Playground tab.)

Benchmark improvement, v2 → v3

Expİvme-DiffusionConversate-v1-Instruct

🧪 Experimental. A 130M-parameter masked-diffusion model, SFT'd for basic chat. Per the model card: output is not reliably grammatical — expect locally plausible word choice that often doesn't compose into coherent sentences. Included here in the spirit of the model card's own honesty about its limitations, not as a working assistant.

İvme-Denoise-v1 (codename Margay 1)

A 2.97M-parameter blind image denoiser (İvmetron-Guanine1, a multi-scale Bayesian shrinkage U-Net). Covers noise levels σ ∈ [5, 50] with one model. It beats DnCNN on every standard benchmark at ~65% of the FLOPs, and sits within 0.22–0.54 dB of DRUNet/Restormer — it is not state of the art on raw PSNR, by design.

Auto estimates a per-pixel noise map from the image itself (best for real photos, where noise is stronger in shadows). Manual lets you set one global σ on the 0–255 scale. Large frames are tiled automatically. Runs fine on CPU, but expect seconds per megapixel.