Sevren Says Embla 1 Runs Offline on a Phone in 1.77 Seconds
Summary
Sevren releases Embla 1, a 4-bit reasoning model it says runs offline on a phone and answers in 1.77 seconds on an iPhone 18 Pro. Sevren reports an average score of 57.4 across five benchmarks, behind five of the twelve models it tested. The weights are on Hugging Face.
Key Points
- Sevren reports Embla 1 averages 888 tokens per answer, the fewest of the reasoning models it tested.
- Embla 1 takes 2.18 seconds in llama.cpp, against 1.77 seconds in Sevren's own Metal implementation.
- Sevren reports Embla 1 scores 76.1 on MATH-500, against 89.6 for Liquid AI's LFM2.5-8B-A1B.