- Updated: April 1, 2026
- 2 min read
Liquid AI Unveils LFM2.5-350M: A Compact LLM Boosting Edge AI with Unmatched Intelligence Density
Liquid AI Unveils LFM2.5-350M: A Compact LLM Boosting Edge AI with Unmatched Intelligence Density
Liquid AI has announced the release of its latest language model, LFM2.5-350M, a 350‑million‑parameter transformer that sets a new benchmark for intelligence density on edge devices. Trained on an unprecedented 28 trillion tokens using a hybrid large‑language‑model architecture and scaled reinforcement learning, the model delivers high‑quality responses while remaining lightweight enough for on‑device inference.
Key Specifications
- Model size: 350 M parameters (≈1.2 GB after quantization)
- Training data: 28 trillion tokens sourced from diverse web corpora
- Training method: Hybrid LIV architecture with scaled reinforcement learning for better alignment
- Benchmark scores: State‑of‑the‑art results on AI agent tasks, outperforming many 1‑B‑parameter models
- Hardware efficiency: Runs at 15 W on a single‑chip ARM Cortex‑A78, achieving >30 tokens/s
Performance Highlights
The LFM2.5‑350M model demonstrates a remarkable intelligence density—delivering comparable accuracy to 2‑B‑parameter models while using less than 20 % of the compute. In benchmark suites such as MMLU and GSM‑8K, it scores 68 % and 71 % respectively, positioning it as a top‑tier compact LLM for edge AI applications.
Edge‑Ready Design
Designed with on‑device deployment in mind, the model can be quantized to 4‑bit precision without significant loss in performance, enabling real‑time inference on smartphones, IoT gateways, and autonomous drones. This makes it ideal for use‑cases like on‑device assistants, low‑latency translation, and privacy‑preserving analytics.
Recommended Use‑Cases
- Smart assistants on wearables and smartphones
- Real‑time language translation in AR/VR headsets
- Edge AI for industrial IoT sensors
- Embedded AI agents for robotics
For a deeper dive into the technical details, read the original announcement on MarkTechPost.
SEO Meta Description
Liquid AI launches LFM2.5-350M, a compact 350M‑parameter LLM delivering high intelligence density for edge AI. Learn specs, benchmarks, and edge‑ready use cases.
Explore related topics: AI Agents, Large Language Models, Edge Computing.
Andrii Bidochko
CTO UBOS
Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.