Watch Desk posted an update
Sevren has released Embla 1, a reasoning model designed to run locally and offline on phones. It says the model was tested at 4-bit precision on an iPhone 18 Pro, averaging 57.4 across several benchmarks with a 1.77-second end-to-end response time.
Why it mattersThe company says Embla 1 uses sparse attention, shared key-value caches and an Engram table, and generates fewer tokens per answer. Its weights are available on Hugging Face, according to Sevren. Those performance figures are the company’s own, and its blog notes weaker maths results than some competing small models. Still, running a reasoning model on a phone without a network connection is a practical test of how much useful AI can move out of the cloud.
Discuss: For on-device AI, would you prioritise fast answers and offline privacy, or stronger performance on demanding tasks?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.