Runway converts Qwen2.5-VL into a diffusion VLM
Runway researchers turned a pretrained autoregressive vision language model into a parallel-decoding diffusion model, a cheaper route to diffusion VLMs that avoids training one from scratch.
Runway Research described A2D-VL 7B (Autoregressive-to-Diffusion) in a post dated September 24, 2025. The model was built by finetuning Alibaba's Qwen2.5-VL for masked block diffusion decoding. The authors are Marianne Arriola, Naveen Venkat, Jonathan Granskog and Anastasis Germanidis.
Diffusion decoders can generate several tokens at once. A confidence threshold sets the balance: a high threshold favours accuracy, and a low one allows more parallel tokens at some cost to quality. Runway says existing diffusion VLMs have four problems. They are expensive to train, since LLaDA-V 8B used at least 12 million visual QA pairs. Their architectures are older than models like Qwen2.5-VL. Their quality drops on long outputs because they diffuse the whole sequence at once. And they only support approximate KV caching.
A2D-VL decodes in blocks of 8 tokens and uses two techniques to keep the base model's abilities during finetuning. Block size annealing starts with small prediction blocks, close to sequential decoding, and grows them to 8 tokens. Noise level annealing first masks the easier tokens nearest the context more often, then moves to uniform masking so the model can generate in any order within a block. Runway's ablations show both techniques help preserve benchmark performance.
According to the post's summary, the conversion needed about 400,000 training pairs, supports exact KV caching, and beat earlier diffusion VLMs on captioning and MMMU-Pro. The source text available here is cut off before the full benchmark results, so detailed scores and speed figures could not be checked.
Runway notes that separate concurrent work also adapted Qwen2.5 for block diffusion, but on text-only tasks.