NVIDIA's quiet Dory 1.0: an open 4B you dial for speed or accuracy
The looped open-weights model picks its compute at serving time — 1 loop for latency, 4 for agentic accuracy — with a 256K window and no hosted endpoint yet.
Copy markdown
One checkpoint, loops on demand
Dory 1.0 is a dense 4B with a looped core — 8 input layers, a 20-layer recurrent block you repeat, then 8 output layers. You set the loop count at inference: 1 for low latency, the default 2, or 4 when an agent needs the accuracy, all off the same weights with no retrain or second download. A 'last-loop' KV-cache mode keeps memory near single-loop size no matter the count.
256K context on a 4B, four ways to run it
NVIDIA shipped four checkpoints — BF16, MXFP8, NVFP4, and a Q5_K_M GGUF — under a commercial-friendly OpenMDW 1.1 license, with a 256K window (262,144 positions validated in vLLM). It's English-only but trained on 43 programming languages across roughly 6T tokens, and on a DGX Spark the NVFP4 build runs about 2.2–2.4x faster than BF16.
The benchmarks, with an asterisk
At 4 loops NVIDIA reports 61.8 on SWE-bench Verified (edging Qwen3.5-9B's 61.0), 73 on AA-LCR ahead of every comparator, and a jump to 14.69 from 5.03 on τ³-bench Banking — looping helps most on agentic and tool-use work. But GPQA and HLE trail the field, every number is from NVIDIA's own harness, and the Hugging Face download counter still reads zero.
Running it today means building from a branch
There's no hosted endpoint and no blog post — NVIDIA just uploaded the weights. vLLM support sits in unmerged PR 60149 and llama.cpp in PR 30306, so serving Dory right now means checking out a branch. The payoff: a single small model you can tune for latency or agentic accuracy per request, testable on one GPU.