VC & PE Glossary
What Is a Diffusion Language Model (DLM)?
Definition
A diffusion language model (DLM) generates text by repeatedly improving multiple positions in an incomplete sequence, rather than predicting one next token strictly from left to right. It can produce many tokens in parallel, but its speed and accuracy depend on the number and cost of refinement steps.
Also called: diffusion LLM · diffusion-based language model · dLLM · DLM
Useful for: Founders, Investors, Developers
A traditional language model writes sequentially: each new token depends on the output it has already produced. A diffusion model takes a different route, beginning with a masked or otherwise incomplete text block and iteratively filling or correcting multiple positions.
Why developers care
Parallel updates can reduce the number of dependent decoding steps. They are especially interesting for code editing, constrained text, agents and fast interactive responses. Yet each refinement pass uses computation; more tokens per second does not automatically mean a lower bill or a better answer.
Real examples
- Inception Mercury 3: a commercially available diffusion model sold through an API. Inception publishes speed claims, while Plotly has measured its performance independently on a data-analysis task set.
- DiffuSpace Clover-35B-A3B-Preview: an open-weight block-diffusion model released in October 2026 with 35B total and about 3B active parameters.
- Google DiffusionGemma: an experimental open 26B MoE model built to explore faster parallel generation.
Diffusion language modeling is not identical to Stanford's Next-Concept Modeling (NCM). NCM learns a hierarchy of discrete representations and predicts each level in parallel. For company funding, product maturity and a fair comparison of speed claims, read VCT's market analysis.
Related terms
- Next-Concept Modeling — Next-Concept Modeling (NCM) is a research framework for generating language through a hierarchy of discrete codes, from coarse representations to finer details. A model predicts multiple codes in parallel within each level, instead of generating text strictly token by token.
Related reading
Editorial note: AI tools assisted with research, structure, or drafting. Venture Capital Tracker retains human editorial responsibility for factual accuracy, relevance, and source quality before publication.