Dory1.0-4B weights are public and expose selectable recurrent depth from one checkpoint.
NVIDIA releases Dory1.0-4B with selectable recurrent depth
2026-10-17Verified announcementTRIAL
NVIDIA released open weights for the 4B-parameter Dory1.0-4B. One checkpoint can run its shared recurrent block one, two or four times, trading more compute for task accuracy. BF16, MXFP8, NVFP4 and Q5_K_M GGUF variants are downloadable with 256K context. Upstream vLLM support remains a pull request, and llama.cpp requires a pinned NVIDIA fork.
What this means for you
For local-model teams, selectable recurrent depth lets one checkpoint cover latency and accuracy tiers, potentially reducing model swaps. Runtime support is not yet in stable upstream releases, and vendor benchmarks do not replace end-to-end tests.
Evidence & related material
Technical details & scores
| Architecture | 8 input layers + shared 20-layer recurrent block + 8 output layers |
|---|---|
| Loops | 1, 2 or 4 from one checkpoint |
| Context | 256K |
| Formats | BF16, MXFP8, NVFP4, Q5_K_M GGUF |
| License | OpenMDW 1.1 |
| Vendor Swe Bench Verified | 50.3 / 59.5 / 61.8 at 1 / 2 / 4 loops |
| Runtime Boundary | vLLM support under review; llama.cpp uses pinned NVIDIA fork |
Technical impact 4
Engineering value 4
Adoption signal 2
Maturity 2
Verification cost 3
Risk 3