Danish startup Sevren has released Embla 1, a reasoning model designed to run directly on a phone — private, offline and free per request. According to the company, the model answers in 1.77 seconds on an iPhone 18 Pro, without sending anything to the cloud.
The name refers to Embla, the first woman in Norse mythology, according to Øresund Startup News — a nod to the model being the first of its kind built by the company.
What Embla 1 is
Embla 1 is a mixture-of-experts model with 10 billion parameters, of which only about 1 billion are active per token. It has 32 layers and was designed from scratch for phone hardware, rather than being a shrunken-down server model. Pretraining used 1.4 trillion tokens, distilling from a larger model of the Qwen family, and training ran on 16 nodes with 8 H100 GPUs each.
To fit on a phone, Sevren used its own sparse attention, shared caches between layers and an “engram” table that avoids recomputing part of the model for every token — techniques detailed in the post on Embla 1's architecture.
Performance: Sevren's own numbers
In the company's tests, at 4 bits, Embla 1 averaged 57.4 across five benchmarks (BFCL, IFBench, AA-Omniscience, GPQA and MATH-500) and answered in 1.77 seconds end to end on the iPhone 18 Pro, using Sevren's own Metal implementation — 2.18 seconds in llama.cpp. The company says the models that scored higher took at least twice as long on the same phone.
These are Sevren's own measurements, on a single phone and with an uncertainty of about ±1 point — with no independent verification yet. And there is a weakness the company itself admits: in math, Embla 1 scored 76.1 on MATH-500, against 89.6 and 90.5 for two small rivals.
On the other hand, among the reasoning models tested, Embla 1 wrote the fewest tokens per answer: 888 on average, against 1,577 to 7,461 for the others — which helps explain its speed.
Availability
Embla 1's 4-bit weights are already available on Hugging Face, for download and local use. Because the model runs on the device, prompts never leave the phone — not even in airplane mode, according to the company.
Read also: Garlic LLM runs a 9-billion-parameter model offline on the PS4.
Reaction
Nordic press highlighted the launch as the most efficient AI model ever built in Denmark, trained on the Gefion supercomputer. On Threads, the Silicon Vikings account shared the announcement, citing Øresund Startup News.
Image credit: Sevren (sevren.ai) — the company's brand card.
Sources
- Sevren — Embla 1: A reasoning model built for phones
- Sevren — The Architecture of Embla 1
- Øresund Startup News — Sevren Releases Embla, Its Most Efficient AI Model That Runs on a Phone
- Silicon Vikings on Threads — the Embla announcement
This article was produced with the help of artificial intelligence, cross-checking public posts and sources on the internet. Please consult the original sources to confirm the information.
By Muse