For those who aren't following the commit history closely, you can see the removal here:
https://github.com/modelscope/ms-swift/commit/a45f1d4f73157ba59062a7fd1f55a40dae759156
This is honestly frustrating because the 35B MoE hits a sweet spot for local deployment. It's powerful enough to handle complex reasoning but small enough to fit on consumer hardware if you've got the VRAM. If the developers are cleaning up the codebase by removing it, it usually implies the model is either being scrapped or pivoted into something else.
I've been trying to build a stable AI workflow around this specific parameter count, and having the rug pulled out via a GitHub commit is a nightmare for planning. Usually, when a model is "coming soon," you see more integration, not deletions.
The impact on local LLM deployment #
The MoE (Mixture of Experts) architecture is exactly why the 35B is so desirable. It gives you the intelligence of a much larger model without the linear increase in compute costs during inference. If this gets shelved, we're left with a gap between the smaller, faster models and the massive ones that require a server rack to run.
I suspect the team might not realize how much the community actually cares about this specific version. Most of the noise happens around the flagship models, but the mid-range MoEs are where the real productivity gains are for individual developers.
If anyone has a lead on whether this is just a temporary reorganization of the codebase or a genuine cancellation, please share. I'm trying to figure out if I should pivot my current prompt engineering efforts to a different model family or keep hoping that this was just a messy merge. It feels like we need to be louder on Hugging Face or X to make sure they know there's an actual demand for the 35B.
Next Why waste RAM on a lightweight Windows box when you can just →