Liquid AI’s LFM2.5-2.6B: Empowering Compact Devices Like Raspberry Pi with Advanced AI—No Cloud or GPUs Required!

This week, Liquid, an AI startup established in 2023 by former computer scientists from MIT, introduced LFM2.5-2.6B, a new open-weight language model tailored for tasks that demand greater agency.

According to the company’s initial announcements and a recent discussion with VentureBeat, LFM2.5-2.6B functions fully on local devices—ranging from smartphones and laptops to Raspberry Pi—eliminating the need for cloud services or GPUs. This development paves the way for edge AI applications and provides enterprises, especially those in regulated sectors, the flexibility to handle sensitive data without risk of cloud exposure.

The model excels in executing high-volume, well-defined tasks directly on user devices, such as automating workflows, tool usage, document management, and background processes. It’s particularly advantageous in settings with limited connectivity, like robotics and vehicles, although more complex coding tasks may be better suited to larger models.

For companies without significant data security concerns, the opportunity to run efficient, task-oriented agents with minimal energy costs may make this model especially attractive.

However, the licensing arrangements for the custom open weights, akin to those of Moonshot’s recent Kimi K3 model, warrant thorough evaluation by enterprise legal teams.

Key Features

LFM2.5-2.6B comprises 2.6 billion parameters, supports a 128,000-token context window, and offers native tool calling. The name combines the model generation (2.5) with its parameter count (2.6B).

Both the fully-trained model and a base checkpoint (LFM2.5-2.6B-Base) for developers interested in fine-tuning are now accessible on Hugging Face. A range of major inference frameworks, including llama.cpp and ONNX, also support the launch, making it versatile for use in consumer devices, enterprise systems, and embedded environments.

In addition, Liquid provides an open-source fine-tuning framework called LEAP.

Rather than positioning LFM2.5-2.6B against larger frontier models, the company argues that a smaller, capable model can effectively open new enterprise applications where latency, confidentiality, deployment versatility, and inference costs hold more significance than sheer performance metrics.

“While I think the top models will be cloud-based, that’s perfectly fine,” said Maxime Labonne, Liquid AI’s head of post-training, in a discussion with VentureBeat. “We aim to cater to a different user base, emphasizing edge AI for situations where cloud models are not feasible.”

Compatible with Minimal Hardware

“A good example is Raspberry Pi,” he noted, citing multiple demos that showcased its efficient operation on such devices.

According to the company, the model achieves decoding throughputs of about 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395, consuming less than 2.5 GB of memory—around 30 tokens per second on smartphones. Users can test the model via Apollo, Liquid AI’s mobile application.

Conversely, on an Nvidia H100 GPU under sustained load, Liquid AI claims that the model can produce nearly 15,000 tokens per second—approximately 1.3 billion tokens daily per card. These benchmarks have not been independently verified.

For Labonne, both memory usage and speed are critical constraints that influence deployability.

“We aim to demonstrate a favorable trade-off, achieving a level of quality comparable to larger models, but within a compact form factor,” he explained. “This enables deployment in scenarios where larger models may not be feasible.”

Designed for Agent Tasks, Not Just Chat

Liquid AI believes that LFM2.5-2.6B was built on the premise that language models are increasingly used within agent frameworks rather than traditional chat interfaces.

“Models are no longer simply used in chatbots; they operate through agentic frameworks like OpenClaw and Hermes Agent,” Labonne noted. “We ensured this model excels not only in math or coding, but also in utilizing various tools.”

The model has been trained on around 34 trillion tokens, features a vocabulary expanded to 128K for enhanced support of non-Latin scripts, and underwent a dedicated mid-training phase to extend its context to 128K tokens for ongoing agentic tasks.

The post-training process involves four stages: supervised fine-tuning, specialized training for distinct domains (such as coding or instruction following), merging expert capabilities into a unified model, and finally, agentic reinforcement learning. During the last phase, training occurred within actual agent frameworks like Hermes Agent and OpenClaw, allowing exposure to real-world tasks like document management and tool invocation.

Labonne described this restructured training approach as a “fortunate accident,” with broader improvements across various capabilities.

“Our new methodologies enhanced our capabilities significantly, even improving areas like code generation—a field where we previously struggled,” he said.

Co-Creating the Model and Harness

Notably, Liquid AI developed its own agent harness to complement the model, showcased in action on phones where it can plan and execute tasks completely on-device.

“This is a harness operating on a phone, and I’m not aware of any other that does this,” Labonne asserted.

Two primary reasons drove this decision: the absence of a phone-native harness and Liquid AI’s interest in a proactive interaction model that anticipates user needs rather than merely responding to prompts.

“We aspire to create proactive agents that monitor user activities and schedule tasks autonomously,” Labonne added. “Such technology is largely absent today.”

The concurrent development of the harness and model allows the software to address the model’s limitations. “The harness compensates for any shortcomings the model has, improving reliability,” he explained. “Ultimately, users care about task completion, not the inner workings of the model versus the harness.”

The model is also compatible with established frameworks like Hermes Agent and OpenClaw, functional behind any OpenAI-compatible endpoint.

Adaptive Reuse for Enterprises

Labonne proposed that this release signifies a transformative potential for small models. Previously, local models primarily served narrowly focused tasks, providing faster and more cost-effective performance than cloud alternatives. Agentic capabilities change this equation, enabling one model to be repurposed simply by adjusting the tools used with it.

“You can transform a calendar assistant into a meeting assistant capable of summarizing discussions—essentially, you modify the tools rather than the model,” he explained. “This kind of flexibility increases generalizability and cost-effectiveness.”

He still advocates for fine-tuning for any production deployments: “Neglecting fine-tuning leaves room for improvement. With proper tuning, it can match the performance of leading models, assuming your task isn’t overly complex,” he remarked, highlighting the accessibility of fine-tuning today.

Competitive Landscape

Liquid AI has released benchmark comparisons showcasing LFM2.5-2.6B against several contenders likely to feature in enterprise considerations: Google’s Gemma models and Alibaba’s Qwen series.

A distinct evaluation by Atomic Chat indicated that LFM2.5-2.6B completed 35 tool calls for three tasks significantly quicker than the DeepSeek-V4-Flash model, which boasts an astonishing 284B parameters but only recently launched in the top tier of OpenRouter.

While Google’s Gemma models are multimodal and designed for various input types, Alibaba’s Qwen series emphasizes scaled reinforcement learning for reasoning tasks. In contrast, LFM2.5-2.6B focuses primarily on text, aiming for efficiency in agentic workflows rather than diluting its capabilities across multiple formats.

Liquid AI’s metrics suggest that its smallest model outperforms the competition in instruction-following and tool-use benchmarks, narrowly trailing only on specific tests compared to larger models.

This competitive edge illustrates the effectiveness of their focused model approach, though licensing varies significantly from alternatives.

Licensing Strategy

LFM2.5-2.6B is distributed under the LFM Open License v1.0, allowing use and modification for organizations earning less than $10 million annually. Larger enterprises must negotiate separate agreements with Liquid AI for commercial usage. Qualified nonprofits are exempt for non-commercial purposes.

Labonne emphasizes the need for a sustainable model development model, balancing commercial licensing with ongoing research contributions to the community.

Asking how compliance would be monitored for larger deployments, he remarked, “That’s more for the legal team, but essentially, we just ask larger companies to reach out to us.”

Liquid AI plans to pair its licensed models with accessible research, including new evaluation methods designed to mitigate common pitfalls in smaller models—issues noted in other models like Qwen.

Enterprise Implications

The launch coincided with a significant partnership announcement from MacPaw, a Ukrainian software firm that will collaborate with Liquid AI to create a localized AI stack for Mac systems.

Liquid AI will provide and refine foundational models for Eney, MacPaw’s macOS assistant, utilizing Apple’s silicon and its own inference engine, with outcomes anticipated later this year.

Labonne highlighted this collaboration as a testament to the practicality and efficiency of smaller models: “Their decision to work with us stems partly from the model’s small size, which aligns with their memory limitations.”

The introduction arrives at a time when hardware makers, operating system developers, and enterprise software firms are heavily investing in local AI capabilities—particularly within ongoing explorations of agent management. Liquid AI’s vision underscores that economic viability, rather than sheer size, will carve out a unique niche in an evolving market landscape.

Ultimately, the significance of small, optimized models for enterprise AI will hinge less on performance benchmarks and more on their operational consistency. Liquid AI’s latest advancements suggest that the focus is shifting from merely achieving larger models to crafting smaller, more efficient ones that fit seamlessly within existing enterprise workflows.