Liquid AI has released a 2.6 billion-parameter model built to move AI agents off cloud servers and onto the devices where users actually work. The LFM2.5-2.6B model, released on August 4, represents a shift in how companies think about deploying artificial intelligence: instead of sending every task to a distant data center, the model runs locally on phones, laptops, and small computers, handling tool calls, data extraction, document management, and background automation without internet dependency.

The founding team behind Liquid AI came from MIT's Computer Science and Artificial Intelligence Laboratory, where they spent years studying liquid neural networks, a type of AI system designed to work efficiently within tight memory and power constraints. CEO Ramin Hasani and his co-founders built the company in 2023 with the conviction that useful AI would need to operate outside data centers, under the real-world limitations of actual devices.

Why Does Running AI Locally on Your Device Matter?

The economics of cloud-based AI create a problem for certain types of work. When an AI agent monitors your calendar, sorts documents, or waits for events in the background, sending every single task to a cloud server means paying for each API call. Those costs add up quickly for repetitive, well-defined work that doesn't require the reasoning power of larger models. Local execution replaces that variable inference bill with hardware and electricity the user or company already controls.

Liquid AI's head of post-training offered a blunt boundary for when to use edge AI: "You should use edge AI when you can't use a cloud model." The company is deliberately leaving frontier-scale coding, broad knowledge work, and the hardest reasoning jobs to larger systems. Instead, LFM2.5-2.6B targets repetitive tasks where sending every prompt, file, and tool result to a remote model adds cost, delay, or privacy risk.

How Does This Model Actually Perform on Real Devices?

Liquid AI reports specific performance metrics across different hardware. The model generates about 220 tokens per second on an Apple M5 Max processor and 113 tokens per second on an AMD Ryzen AI Max+ 395, while using less than 2.5 gigabytes of memory. On smartphones, the company reports roughly 30 tokens per second, and the model runs effectively on Raspberry Pi-class hardware, the type of small single-board computers used in embedded systems. These measurements have not been independently verified, and exact results will depend on quantization methods, specific hardware, context length, and workload.

The model includes native tool calling, meaning it can select and use external software functions without additional prompting. It also features a 131,072-token context window, allowing it to process roughly 100,000 words at once, far larger than typical smartphone applications require.

What Training Process Makes This Model Work for Agents?

Liquid AI used a four-stage training process that directly addresses a practical problem: a model can score well on isolated question-answering tests and still fail when it must follow instructions, select a tool, format arguments correctly, read results, and decide what to do next. The company trained the model inside actual agent software, including Hermes Agent and OpenClaw, handling tasks that required research, coding, document management, and external tool use.

• Pretraining: Liquid AI pretrained LFM2.5-2.6B on roughly 34 trillion tokens, doubled the vocabulary to 128,000 tokens, and added a dedicated context-extension phase to handle longer documents and conversations
• Supervised Fine-Tuning: The model learned from human-labeled examples showing correct tool use and task completion patterns
• Specialist Teacher Models: Liquid AI used larger models to guide the smaller model's learning, a technique called distillation that transfers knowledge from bigger systems to smaller ones
• Agentic Reinforcement Learning: The final stage trained the model to complete multi-step tasks by rewarding successful tool selection and task completion sequences

This approach directly addresses why many small models fail in production: they haven't practiced the full sequence of following a system prompt, selecting a tool, formatting arguments, reading results, and deciding what to do next.

How Does This Compare to Other Small Models?

Liquid AI's own evaluations show LFM2.5-2.6B competing with larger Gemma and Qwen models across instruction following, tool use, and several agent benchmarks. The results also show the boundary the company acknowledges: larger models retain an advantage on demanding coding work. However, these comparisons remain vendor-run tests, so developers will need to measure tool-selection accuracy and multi-step reliability against their own workloads before committing to the model.

The model ships with support for multiple inference frameworks, including llama.cpp, MLX, vLLM, SGLang, and ONNX, giving developers flexibility in how they deploy it. Liquid AI also released a base checkpoint for developers who want to fine-tune the model for their specific use cases.

What Are the Commercial Terms and Licensing?

Developers can download and modify LFM2.5-2.6B, although Liquid AI's terms differ from a standard permissive open-source license. Under the LFM Open License, free commercial rights end when a user's legal entity reaches $10 million in annual revenue. Businesses above that threshold must purchase a commercial license from Liquid AI. This provision gives individual developers and smaller startups wide access while preserving an enterprise licensing business for the company.

That licensing boundary creates an early procurement question for companies evaluating the model against alternatives distributed under more permissive terms. Mistral's Ministral 3 3B, for example, carries Apache 2.0 terms and offers a 256,000-token context window, providing a different licensing path for some organizations.

Where Is Liquid AI Deploying This Technology?

Liquid AI has already tied its device-native thesis to larger deployment partners. Mercedes-Benz announced a multi-year partnership to develop embedded intelligence for in-car voice and language systems, where local processing reduces latency and protects driver privacy. Shopify's agreement with Liquid AI covers low-latency models for search and other commerce workloads, where millisecond response times matter for user experience.

"You should use edge AI when you can't use a cloud model," stated Maxime Labonne, head of post-training at Liquid AI.

Maxime Labonne, Head of Post-Training at Liquid AI

Liquid AI has substantial capital backing the effort. The company announced a $250 million Series A in December 2024, led by AMD, after disclosing $46.6 million in seed financing in 2023. The Series A was intended to fund compute, model development, and deployment tooling for edge and on-premise customers.

What Does This Mean for the Future of AI Deployment?

The release extends a product line that is filling out quickly. Liquid AI shipped LFM2.5-8B-A1B, an 8 billion-parameter mixture-of-experts model with 1.5 billion active parameters and a 128,000-token context window, in May. The new 2.6 billion-parameter release moves further toward phones and embedded computers while retaining the tool-calling focus that makes agents practical.

Hasani's bet is that deployment performance will carry enough value to support the licensing boundary. If LFM2.5-2.6B can reliably operate tools on low-cost processors, local agents become viable in vehicles, regulated workplaces, and consumer devices where a permanent cloud connection is expensive, undesirable, or unavailable. The release gives developers the model weights and tooling to test that claim. The harder proof will come from agents that continue completing real work after the demo ends.