News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnLFM2

LFM2

Hybrid language models for edge and on-device use

Liquid AI released LFM2, a series of small hybrid language models designed for efficient deployment on smartphones and laptops. The family includes four sizes trained on ten trillion tokens, offering fast inference on central processing units while supporting multiple languages and tool use.

Share on WhatsApp

  • Trained on 10 trillion tokens across four model sizes
  • Hybrid architecture uses convolutions and attention layers
  • Supports tool use via JSON and Python function calls
  • Runs efficiently on CPU, GPU, and NPU hardware
  • Context window extends to 32,768 tokens

What it is

LFM2 is a family of compact language models built by Liquid AI for edge computing and on-device applications. The series comprises four checkpoints with parameter counts ranging from 350 million to 2.6 billion. These models aim to balance speed, memory efficiency, and performance for tasks such as data extraction and multi-turn conversations. They are not intended for heavy programming or knowledge-intensive work without further tuning.

How it came about

Liquid AI developed LFM2 as the second generation of its foundation models. The team used knowledge distillation from the earlier LFM1-7B model to train these smaller variants. The release includes post-trained checkpoints designed to improve upon the previous generation in training speed and inference efficiency on standard hardware.

Inside the model

The architecture combines short-range convolution blocks with grouped query attention layers, using multiplicative gates. Each model was trained on approximately ten trillion tokens, consisting mostly of English text, alongside multilingual and code data. The training process involved supervised fine-tuning and direct preference optimization with length normalization. The context window supports up to 32,768 tokens.

What it can do

The models support eight languages, including English, Chinese, and Arabic. They can execute tools via JSON definitions and Pythonic function calls. Benchmarks show strong results in instruction following and mathematics compared to similar-sized competitors. For instance, the 1.2B variant scores above 55 on MMLU and nearly 58 on GSM8K, outperforming several larger open models in specific categories.

How to use it

Weights for all four sizes are available under the LFM Open License v1.0. Users can run the models using Hugging Face Transformers, vLLM, or llama.cpp with GGUF formats. The developer recommends fine-tuning the models on specific narrow use cases to achieve optimal performance for agentic tasks or creative writing.

Questions people ask about LFM2

What hardware does LFM2 run on?

LFM2 is designed for flexible deployment on central processing units, graphics processing units, and neural processing units. It works well on smartphones, laptops, and vehicles due to its memory efficiency and fast decode speeds.

Which languages does LFM2 support?

The model supports eight languages: English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish. The training data mixture was approximately 75 per cent English, 20 per cent multilingual, and five per cent code.

Can LFM2 write computer code?

The developers do not recommend using LFM2 for tasks requiring programming skills. While it was trained on some code data, it is better suited for agentic tasks, data extraction, and general conversation after fine-tuning.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on LFM2 yet, so the facts on this page come from the sources listed above: huggingface.co/LiquidAI/LFM2-1.2B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.