Extend the complete hybrid drafter to other model families

Develop a mobile-executable native draft head for other model families so that the complete HyCo hybrid drafter, including both the native head and the SLM-based alternative-branch drafter, can operate on the device.

Background

The cross-family experiments show that HyCo’s SLM Drafter can be transferred to Muse-Glimmer-30B through a token bridge, but the verifier’s DFlash2 head remains on the server because it was not deployed on the device. The paper therefore identifies the unresolved task of enabling a suitable native draft head on the mobile device for other model families, which would allow the entire hybrid drafting process—not merely its SLM component—to move to the device.

References

A head that the mobile device can run for such families would move the whole Hybrid Drafter, not only its SLM half, to the device, which we leave to future work.

— HyCo: Hybrid Drafting for Cost-Efficient Mobile-Cloud Speculative Decoding  (2610.21436 - Bae et al., 15 Oct 2026) in Appendix, Section “Additional Cross-Family Measurements” (Appendix \ref{app:crossfamily})