Open source #Open models Oct 18, 2026 · Originally published Oct 16

NVIDIA releases Dory1.0, a looped 4B model that trades compute for accuracy by loop count: SWE-bench Verified goes from 50.3 at one loop to 61.8 at four

One checkpoint, several compute budgets, with no model swap or retraining. What a looped model is and how much accuracy the loops buy.

Primary source NVIDIA on Hugging Face · huggingface.co Read the original ↗

// Key points

4B parameters, a looped (latent-recurrent) normalized-GPT with interleaved sliding-window and full-attention layers, up to 256K context, English only.
Compute is set by how many times the shared block runs: SWE-bench Verified is 50.3, 59.5 and 61.8 at 1, 2 and 4 loops, against 61 for Qwen3.5-9B in the same table; NVIDIA says more loops solve the same tasks with fewer generated tokens.
Last-loop KV caching keeps one cache per recurrent attention layer, so memory doesn't grow with each added loop.
Pre-trained natively in NVFP4, offered in BF16, MXFP8, NVFP4 and Q5_K_M, under the OpenMDW-1.1 license.

Builder's takeOne model with compute set per request suits on-device and small teams: one loop for easy requests, four for hard ones, without maintaining two models. I'd measure the real loop-versus-latency curve in an offline batch job first; it's English-only, so Chinese products shouldn't ship it yet.

// Background · from #Open models

Full timeline →