How to use DSpark with OpenVINO

DSpark is not supported by OpenVINO GenAI

You cannot use DSpark with OpenVINO GenAI.

OpenVINO GenAI supports speculative decoding, but DSpark is not currently exposed as a supported speculative decoding method in OpenVINO GenAI. A DSpark GGUF therefore cannot be loaded as a DSpark draft model alongside a main GGUF through the GenAI API.

If you are trying to run a main GGUF with a DSpark GGUF as its draft model, OpenVINO GenAI does not currently support that setup.

DSpark support in progress

No DSpark issue exists in the OpenVINO GenAI, OpenVINO, or optimum-intel trackers. The work is in pull requests.

Add DSpark speculative decoding (preview) [part of minicpm-v-4.7 support] is open and unmerged. It was opened 17 October 2026. It adds DSpark on top of the DFlash pipeline as an experimental preview, loaded through the existing draft_model() API into LLMPipeline and VLMPipeline on CPU and GPU. The preview is greedy only, single request, and without prefix caching. No release version is named in the pull request. It is to be merged after its companion export pull request, and one reviewer approved it on 27 October 2026 with other code owner reviews still outstanding.

[DSpark] Export DSpark drafts with a vanilla Markov head [part of MiniCPM-V 4.7 support] is also open and unmerged. It was opened 13 October 2026, and a reviewer requested changes on the same day.

The scope is narrow. The pull requests cover MiniCPM-V 4.7 drafts in DeepSpec format, exported with optimum-cli to OpenVINO IR. They do not cover speculators-format DSpark drafts, which fail when loaded under this work. The described path is an exported OpenVINO IR draft folder, not a raw DSpark GGUF file passed to the API.

Conclusion: DSpark is not supported in a released OpenVINO GenAI build yet. Support is in progress as an unmerged preview for MiniCPM-V 4.7, with no delivery version named.