Drex DLM
Nace.AI's decision model on NVIDIA's Efficient-DLM-8B diffusion backbone: the state is read bidirectionally, each question branch causally, and a pointer head returns a probability for every option in one forward pass. Also sold as the hosted Drex API.
Nace merged a trained decision adapter into Efficient-DLM-8B and added a pointer head; question branches cannot see each other. The weights take 32,768 tokens; the local runners default to 16,384. It runs through a Python server and Nace's own llama.cpp and Ollama forks; a Q8_0 GGUF sits in a sibling repo. Only tested on Apple Silicon, and long-context quality is experimental; the repo's own 6 October cross-runner check is marked not signed off. The card says choice and score confidence is not a calibrated probability of correctness. Nace quotes 52.31 on the third-party Decision Index 0.2 from its own run; Drex was not on the public board as of 6 October. Weights are CC BY-NC 4.0 from the NVIDIA base, so no commercial rights and no mirroring; code is MIT. The hosted Drex API (drex-v1.0 from 24 September, v1.5 from 28 September) is paid per input token, and Nace does not say which hosted version the open weights match.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 8.2B |
| Base model | nvidia/Efficient-DLM-8B |
| Maker | Nace.AI |
| Released | 2026-10-01 |
| License | cc-by-nc-4.0 |
Hosted API
Served by Nace.AI — $0.05/MTok input, $0/MTok output. Get access · API docs.
Get the weights
pip install systemonemodels
systemone pull nace-ai/drex-dlm
The files are served from the maker's Hugging Face repository, nace-ai/drex-dlm, and verified against the checksums recorded here.