MDLM-en-1.7B-SFT
MDLM-en-1.7B-SFT is an English instruction-tuned masked diffusion language model. Its bidirectional transformer generates text through iterative denoising rather than ordinary left-to-right decoding.
Research release
Model design and training
The card names LumiOpen/mdlm-en-1.7b as its parent and describes pretraining on ten billion FineWeb tokens. SFT uses English Dolci instruction data, followed by a short synthetic conversation stage. Its sampler uses loglinear SUBS noise scheduling and top-p filtering. The named parent was not separately available in the reviewed catalog.mdlm-en-1.7b-sft: official model card
Practical assessment
Practical guidance. This is a research option for comparing generation methods. Measure a complete answer, not just throughput on a full output block. The card reports different speed behavior for long generation and short chat replies; inspect coherence and repetition as well as elapsed time.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | mdlm-en-1.7b-sft |
| Model size | 1.7 billion parameters.mdlm-en-1.7b-sft: official model card |
| License | Apache 2.0mdlm-en-1.7b-sft: official model card |
| Sampling | Iterative masked-token denoising; custom generation function.mdlm-en-1.7b-sft: official model card |