Developer tests Qwen 35B MoE model running locally on Galaxy S26 Ultra
A developer ran an early experiment getting a 35-billion-parameter Qwen model to run directly on a Samsung Galaxy S26 Ultra phone. Because MoE models only activate part of the network at a time, the active portion fit within the phone's memory limits. The exact methods and used were not disclosed.
After optimization, the setup reportedly achieved roughly 90 tokens/second for input processing and about 8 tokens/second for output generation on the device. The author says they learned AI/ML independently without a formal degree, and has compute and resources to keep testing. They also mention four papers submitted to arXiv that remain on hold for review because they are first author without an academic institution affiliation.
They're looking for collaborators to test further.
Key points
- Ran a ~35B-parameter Qwen MoE model locally on a Galaxy S26 Ultra phone
- Input processing around 90 tokens/second, generation around 8 tokens/second
- method and not disclosed
- Self-taught author with no formal degree; 4 arXiv papers on hold due to lack of institutional affiliation
- Looking for collaborators to continue testing