Thursday, 19 November 2026, story 5 of 7
Hugging Face shows 180B MoE model running in 32GB of RAM
The post pairs Transformers with expert streaming and a demo Space, but the text available gives no speed figures or setup details.

What happened
Hugging Face has published a blog post titled "180B Parameters, 32GB of RAM: Pushing MoE Inference to the Limit with Transformers". By its title, the post covers running a 180B-parameter mixture-of-experts model in 32GB of RAM using the Transformers library. It links a Space called "MoE expert streaming in Transformers", described as generating text with Qwen3.8 using streamed experts.
The page text lists six models, with entries of 180B, 177B, 36B, 28B and 7B parameters. It does not say which one was run. It also does not say whether the feature is in a released Transformers version, how fast generation is, or what it costs to run.
What it means
Streaming experts suggests that only part of a mixture-of-experts model's weights sit in RAM at any moment, with the rest fetched as needed. If so, memory capacity stops being the hard limit on local hardware, and storage and transfer speed likely take its place. The source does not confirm the mechanism.
Fitting in 32GB is not the same as being usable. If generation is slow, the approach suits offline or batch work better than interactive coding assistants. The figures that decide this are tokens per second, disk requirements and the hardware used, and none appear in the available text. Developers should read those numbers in the full post before treating a 180B model as a laptop option.
What to do: Before planning around it, check the post's speed and storage figures and try the Space; fitting in 32GB says little about usable speed.
Source
Filed under Open source. Published .
"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.