Dory 1.0 4B
NVIDIA's 4B open looped model whose compute is set at serving time via loop count (1, 2, or 4), with 256K context.
This entry is marked for review. Check the original sources before relying on its details.
Overview
NVIDIA released Dory 1.0 4B around October 16, 2026 as a 4B-parameter open-weight model that reuses a shared block of layers, so developers change inference depth by setting the loop count at serving time without retraining or swapping checkpoints. Secondary write-ups describe an 8-layer input block, a shared 20-layer recurrent block, an 8-layer output block, last-loop KV caching, and a 256K-token context, with 2 loops as the recommended default. NVIDIA-reported results are 61.8% on SWE-bench Verified at four loops versus 50.3% at one, and 22.2% versus 12.5% on Terminal-Bench 2.1. No third party has confirmed the numbers, and the model is reportedly English-only.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Looped architecture: loop count (1, 2, or 4 benchmarked) chosen at serving time
- ~4B parameters with 256K-token context; last-loop KV caching keeps memory roughly constant as loops rise
- Vendor-reported SWE-bench Verified 50.3% (1 loop) to 61.8% (4 loops)
- Weights reportedly in BF16, MXFP8, NVFP4, and GGUF under the OpenMDW-1.1 license
- vLLM and llama.cpp support reportedly required open pull requests at launch
Use cases
- Trading latency for accuracy per request with one checkpoint
- Small-footprint coding and agent workloads on local GPUs
Pricing
Free open weights (OpenMDW-1.1 license per secondary sources).
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 16 Oct 2026
- Record updated
- 23 Oct 2026
- Record last checked
- 23 Oct 2026