Danish company Sevren releases Embla 1, a reasoning model designed for phones
Sevren says Embla 1 runs on a phone without a network connection. In the company’s own comparison, five other models scored higher, but they answered more slowly on an iPhone 18 Pro.
Copenhagen-based AI company Sevren introduced Embla 1 in a blog post on 13 November. Embla 1 is a reasoning model: a language model that writes intermediate steps before giving its answer. Sevren set out to build a fast model that uses few tokens per answer.
Larger Qwen models acted as teachers
Embla 1 is a mixture-of-experts (MoE) model: each token passes through 8 of 96 experts plus one expert shared by all tokens. Some layers compute attention weights with a ReLU² function instead of softmax. Sevren said about 95 per cent of the attention weights are then exactly zero, and the model skips them when generating text. Separate bridge modules build the cache for the second half of the model, so the model skips layers 16–31 when processing the prompt.
Sevren trained the model on 16 compute nodes, each with eight H100 GPUs. In every stage before GRPO reinforcement learning, Sevren taught Embla to imitate the predictions of a larger model. Pretraining used 1.4 trillion tokens, with Qwen3.5-35B-A3B-Base as the teacher. For the last 2,000 fine-tuning steps, Sevren trained the model in the 4-bit format it uses on the phone (quantisation-aware training).
During reinforcement learning, automated checkers scored answers in areas such as maths and code. On knowledge questions, a wrong answer cost a point and declining to answer scored zero. In the final stage, the model was penalised for repeated sequences. Sevren said the share of answers that ran to the 16K-token limit fell from 20.7 per cent to 1.8 per cent.
Higher-scoring models were slower
Sevren compared Embla with 12 small models, all run at 4 bits. The score is the average of five benchmarks: BFCL, IFBench, AA-Omniscience, GPQA and MATH-500. Latency is the response time on an iPhone 18 Pro with a 1,024-token prompt and 256 generated tokens.
The table shows Embla and seven other reasoning models. Embla’s speed is not directly comparable with the others’, because Sevren ran Embla with its own Metal implementation and the other models with llama.cpp. With llama.cpp, Embla’s latency was 2.18 seconds.
| Model | Score | Latency (s) | Tokens/answer |
|---|---|---|---|
| Embla 1 | 57.4 | 1.77 | 888 |
| LFM2.5-8B-A1B | 58.8 | 3.59 | 3,129 |
| LFM2.5-2.6B | 62.7 | 5.33 | 3,557 |
| Granite 4.2 3B | 61.0 | 6.80 | 5,633 |
| Gemma 4 E4B | 60.4 | 9.53 | 1,577 |
| Qwen3.5-4B | 60.0 | 9.69 | 6,619 |
| Gemma 4 E2B | 52.8 | 4.95 | 1,611 |
| MiniCPM5-1B | 43.1 | 1.95 | 6,461 |
Source: Sevren’s own measurements, 4-bit models, up to 16K tokens per answer.
Sevren points out that the figures come from one phone, one evaluation program and five benchmarks. The averages carry an uncertainty of about ±1 point, and the ranking may change under other conditions. LFM2.5-8B-A1B, for example, leads by 1.4 points. Embla falls clearly behind in maths, the company said: it scored 76.1 on MATH-500, against 89.6 for LFM2.5-8B-A1B and 90.5 for LFM2.5-2.6B.
Embla’s 4-bit weights can be downloaded from Hugging Face, and Sevren says queries on the phone cost nothing. The company has published embla-metal, a runtime that uses Apple’s Metal API, on its GitHub page.
Read more: Models