The Qwen 35B MoE disappeared from recent ms-swift commits
I was reviewing the latest changes in the ms-swift repository and spotted something that feels like a major warning sign for anyone depending on that particular model size. The Qwen 35B MoE has been removed from newer commits, and it's beginning to look like the model may never receive a full official release.
What was removed from the ms-swift repository?
For those not tracking the commit history closely, the removal appears here:
https://github.com/modelscope/ms-swift/commit/a45f1d4f73157ba59062a7fd1f55a40dae759156
Why is the 35B MoE critical for local deployment?
This is genuinely frustrating because the 35B MoE occupies a sweet spot for local deployment. It delivers enough power for complex reasoning while staying small enough to run on consumer hardware with sufficient VRAM. When developers scrub a model from the codebase, it typically signals the project is either being abandoned or redirected.
I've been constructing a stable AI workflow around this exact parameter count, and losing it through a GitHub commit derails planning entirely. Normally, an upcoming model shows increased integration, not deletions.
The impact on local LLM deployment
The Mixture of Experts architecture is precisely why the 35B is so attractive. It provides the intelligence of a much larger model without the linear compute cost increase during inference. If this gets shelved, we're left with a gap between smaller, faster models and massive ones that demand server-grade infrastructure.
Does the team understand the community's needs?
I suspect the team may not grasp how much the community values this specific version. Most attention focuses on flagship models, but mid-range MoEs deliver the real productivity gains for individual developers.
If anyone knows whether this is a temporary codebase reorganization or a genuine cancellation, please share. I'm deciding whether to redirect my prompt engineering work to another model family or hold out hope this was just a messy merge. It feels like we need to be more vocal on Hugging Face or X to demonstrate actual demand for the 35B.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
I’d love to clarify this—based on the documentation from the SWIFT project, it appears they explicitly integrated the MoE (Mixture of Experts) layer into the larger model architecture rather than discarding the weights entirely. The paper and docs suggest a seamless reintegration where the specialized components retain their trained parameters within the expanded framework.
Manually pinning my version was the only way to avoid the crash, especially since I noticed the model weights in the latest commit were missing—double-checking the ModelScope documentation confirmed the change before reverting. Did the commit actually delete it?
Had to roll back my environment after the Qwen 35B disappearance—anyone else seeing pipeline crashes? For now, I’ve switched to SWIFT’s lightweight fine-tuning infrastructure (docs here) to avoid dependency issues, though I’m still debugging the root cause.