Liquid AI · 2026-05-28 · major
Liquid AI Ships LFM2.5-8B-A1B — On-Device MoE With 1B Active Parameters, 38T Training Tokens, 128K Context, and Explicit Chain-of-Thought
Hybrid MoE with 8B total / 1B active parameters trained on 38T tokens, 128K context, doubled vocab, and an explicit reasoning trace — pitched as the fastest on-device assistant in its size class on CPU and GPU.
8B-total / 1B-active mixture-of-experts trained on 38 trillion tokens, tuned for on-device agentic work with explicit chain-of-thought.
Key specs
| Parameters | 8B |
|---|---|
| Active params | 1B |
| Context window | 128K |
| Training tokens | 38T |
| Vocab size | 128K |
| Ifeval | 91.84 |
| Math500 | 88.76 |
| Aime25 | 42.53 |
| Tau2 telecom | 88.07 |
What is it?
LFM2.5-8B-A1B is Liquid AI's new on-device assistant model. It is a hybrid mixture-of-experts that only activates 1B parameters per token, so it stays fast on phones and laptops, but it was pretrained on the same scale of data as much larger frontier models.
How does it work?
The architecture combines MoE routing with grouped-query attention and a gated short-convolution path inherited from the LFM2 family. Pretraining was scaled from 12T tokens (LFM2) to 38T tokens, the context window jumps from 32K to 128K, and the vocabulary doubles to 128K to lift tokenization efficiency by 120%+ in Hindi, Thai, Vietnamese, Indonesian, and Arabic. The model emits an explicit chain-of-thought before answering and is post-trained with reinforcement learning for tool-call chaining.
Why does it matter?
Most on-device models trade reasoning quality for speed. Liquid's IFEval (91.84), MATH500 (88.76), and Tau² Telecom (88.07) scores put LFM2.5-8B-A1B in range of much larger dense models while keeping the latency budget consumer-friendly, which makes it usable inside apps that cannot ship cloud inference.
Who is it for?
edge / mobile app developers and indie tinkerers
Try it
huggingface-cli download LiquidAI/LFM2.5-8B-A1B