Liquid AI Unveils LFM2: Claimed to be the Fastest On-Device Foundation Model, Combining Performance and Speed

The startup Liquid AI has launched its second-generation foundation model, LFM2, designed specifically for edge devices like phones, laptops, and AI PCs. This article delves into the three models of LFM2, their impressive performance benchmarks, a comparison with models like Qwen 3 and Llama 3.2, and analyzes the significance of its open-source release for developers and the industry.

The development of artificial intelligence is rapidly moving from the cloud to the various devices in our daily lives. In line with this trend, the startup Liquid AI recently dropped a bombshell with the official release of its second-generation Liquid foundation model series—LFM2. This series of models is not only touted as the fastest and most memory-efficient “on-device AI” on the market but also comes with its model weights open-sourced, beckoning to the global developer community.

LFM2’s goal is crystal clear: to enable AI to run seamlessly and efficiently on phones, laptops, AI PCs, cars, and even wearable devices and robots. This means that the generative AI we experience in the future will no longer be entirely dependent on remote servers but will be able to react in real-time locally, providing an unprecedentedly smooth experience.

The LFM2 Series Models: Three Compact and Powerful Choices

This time, Liquid AI has launched three LFM2 models of different scales at once: LFM2-350M, LFM2-700M, and LFM2-1.2B. The parameter scales of these models range from 350 million to 1.2 billion, designed to meet the computing capabilities and application needs of different devices.

• LFM2-350M: This is the most compact model in the series, designed for devices with extremely limited resources, such as smartwatches or simple IoT devices.
• LFM2-700M: As the middleweight contender, it strikes an excellent balance between performance and speed, making it ideal for smartphones and most laptops.
• LFM2-1.2B: This is currently the most powerful model in the series, targeting high-end AI PCs and edge computing scenarios that require more complex reasoning capabilities.

Interestingly, Liquid AI’s chart shows a neat linear growth relationship between the model size and performance score of the LFM2 series. This means users can clearly expect that choosing a larger model will result in stronger processing power.

Performance Showdown: How Does LFM2 Dominate the Field?

Actions speak louder than words, and Liquid AI has provided detailed performance test data, pitting LFM2 against several mainstream lightweight models on the market, including Alibaba’s Qwen 3, Meta’s Llama 3.2, and Google’s Gemma 3.

Based on the officially released data, LFM2’s performance is indeed impressive.

In several industry-recognized benchmarks, the LFM2-1.2B model demonstrated its superiority in many areas:

• MMLU (Massive Multitask Language Understanding): LFM2-1.2B scored a high of 55.23, surpassing Llama 3.2-1B-Instruct’s 46.6 and Gemma 3-1B-it’s 40.08.
• GSM8K (Grade School Math): In this test of logical reasoning, LFM2-1.2B took a significant lead with a score of 58.3, showcasing its excellent mathematical problem-solving abilities.
• IFEval (Instruction Following Evaluation): LFM2-1.2B achieved a score of 74.89, closely following Qwen 3-1.7B’s 73.98, a very commendable performance.

These figures show that even though LFM2 has fewer model parameters than its competitors, its overall capabilities are comparable, and in some aspects, even superior. This is what Liquid AI refers to as “high efficiency”—doing more with less.

Speed is King: The On-Device Generation Experience

For on-device applications, the model’s inference speed directly impacts the user experience. Imagine giving a command to your phone and having to wait several seconds for a response—that would surely be frustrating.

Liquid AI has particularly emphasized the astonishing text generation speed of LFM2. In a chart showing tests conducted on an AMD H8370 CPU, the generation speed of the LFM2 models (in tokens/sec) at different context lengths far outstrips that of its rivals.

Especially LFM2-350M and LFM2-700M, their speed remains at a very high level even when processing longer texts. This means that whether for real-time Q&A, drafting text, or code assistance, LFM2 can provide an extremely smooth, almost latency-free interactive experience. This is undoubtedly a huge draw for the AI PC and high-end smartphone markets that pursue the ultimate user experience.

The Significance of Open Source: Accelerating the Adoption of Edge AI

This time, Liquid AI is not just launching a product; it is also open-sourcing the model weights of LFM2 for both academic research and commercial use. This decision is of great significance to the entire AI community.

Open-sourcing allows developers to freely customize and fine-tune LFM2 to build applications that meet specific needs. For example, smartphone manufacturers can use LFM2 to develop smarter voice assistants, while car companies can create more responsive in-vehicle infotainment systems.

By harnessing the power of the open community, the LFM2 ecosystem can grow rapidly, sparking more innovative applications and truly accelerating the popularization of edge AI technology.

Frequently Asked Questions (FAQ)

Q: How is the LFM2 model different from other large language models (like GPT-4)?

A: The main difference lies in their design goals and application scenarios. Large models like GPT-4 primarily run on cloud servers, aiming for ultimate comprehensive capabilities. LFM2, on the other hand, is specifically designed for “edge devices” or “on-device” applications. It focuses more on speed and efficiency, with the goal of providing the best real-time AI experience on limited hardware resources.

Q: As a developer, how can I start using LFM2?

A: Liquid AI has open-sourced the model weights for LFM2. You can visit their official X (formerly Twitter) account @LiquidAI_ to find the relevant open-source links and technical documentation to download the model and start your development project.

Q: What does “Liquid” in LFM2 mean?

A: The word “Liquid” comes from the “Liquid Neural Networks” developed by Liquid AI. This is a network architecture inspired by biological nervous systems. Compared to traditional static neural networks, it can adapt more dynamically to changes in input, giving it a unique advantage in processing time-series data and improving computational efficiency. This is one of the key reasons why LFM2 is so efficient.

In conclusion, Liquid AI’s LFM2 series has undoubtedly set a new benchmark for edge computing AI. It not only proves that small models can possess powerful performance but also, through its open-source approach, invites developers worldwide to jointly shape a more responsive and personalized AI future.