Sevren releases Embla 1, an offline phone inference model
·2026.11.18 08:30
Key point
Sevren released the offline phone inference model Embla 1 under the MIT license.
Details
- Embla 1 model with a 10B-A1B MoE structure released, applying the MIT license
- Recorded a latency of 1.77 seconds when processing a 1,024-token prompt and generating 256 tokens on the iPhone 18 Pro
- Achieved an average score of 57.4 across 5 benchmarks, confirming speeds 2.0 to 5.5 times faster than top comparable models
- Provides the embla-metal runtime exclusively for Apple Silicon, with generation speeds of 290 to 339 tok/s on the M5 Pro
- Confirmed inefficient Korean tokenization, using approximately 4.4 times more tokens compared to Qwen3.8
- Supports 256K context, applying Sevren Sparse Attention and KV bridge technology
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.