[Roadmap] DeepSpeed Roadmap Q1 2026

[PKUWZP]

This is a living document! For each item here, we intend to link the RFC as well as discussion Slack channel in the DeepSpeed Slack.

New Accelerator Support

• [ ] DeepSpeed support on TPU

Emergent Model Architectures

• [ ] SuperOffloading for Mixture-of-Expert (MoE) Training

Reinforcement Learning

• [ ] DeepSpeed backend integration as the training engine for verl

New Optimizer Support

• [ ] Muon Optimizer Support for ZeRO3

[delock]

Does "DeepSpeed backend integration as the training engine for verl" means to be the default training engine for verl?

[sfc-gh-truwase]

Does "DeepSpeed backend integration as the training engine for verl" means to be the default training engine for verl?

No, I believe this means integrating DeepSpeed as one of the backend options alongside FSDP and Megatron.