Liquid AI ships a 2.6B model for agents that stay on-device

Why it matters

Liquid AI is turning on-device AI into an agent deployment strategy: lower recurring inference costs, local data handling and offline operation, balanced against unverified benchmarks and a commercial license threshold.

Ramin Hasani, the co-founder and CEO of Liquid AI, released a 2.6 billion-parameter model on August 4 built to move a specific class of AI agents off cloud servers and onto the phones, laptops and small computers where their work originates.

The LFM2.5-2.6B release is the clearest expression yet of the argument Hasani has pursued since starting Liquid AI in 2023 with Mathias Lechner, Alexander Amini and Daniela Rus: useful AI will need to run outside data centers, under the memory, latency and power constraints of real devices.

That founding group came out of MIT's Computer Science and Artificial Intelligence Laboratory, where its work on liquid neural networks focused on building smaller computational systems for dynamic environments. Hasani, who earned his doctorate in computer science from TU Wien, previously worked as a principal AI and machine-learning scientist at Vanguard and conducted postdoctoral research with Rus at MIT CSAIL.

LFM2.5-2.6B packages that research agenda for the agent market. Liquid AI designed the model for tool calls, data extraction, document handling, workflow automation and long-running background tasks. The model has native tool calling and a 131,072-token context window, according to its Hugging Face model card. Liquid AI also released a base checkpoint for developers who want to fine-tune it.

In an August 6 report, Liquid AI head of post-training Maxime Labonne gave the product strategy a blunt boundary: "You should use edge AI when you can't use a cloud model."

That constraint-first pitch is important. Liquid AI is leaving frontier-scale coding, broad knowledge work and the hardest reasoning jobs to larger systems. LFM2.5-2.6B is aimed at repetitive, well-defined work where sending every prompt, file and tool result to a remote model adds cost, delay or privacy risk.

Training the model inside agent software

Liquid AI says it pretrained LFM2.5-2.6B on roughly 34 trillion tokens, doubled the vocabulary to 128,000 tokens and added a dedicated context-extension phase. Those are Liquid AI's figures and training descriptions.

The more consequential change came after pretraining. Liquid AI used a four-stage process covering supervised fine-tuning, specialist teacher models, multi-domain on-policy distillation and agentic reinforcement learning. During the final stage, LFM2.5-2.6B operated inside agent harnesses including Hermes Agent and OpenClaw, handling tasks that required research, coding, document management and external tools.

Training inside the software that will consume the model addresses a practical problem for small models. A model can score well on isolated question-answering tests and still fail when it must follow a system prompt, select a tool, format an argument correctly, read the result and decide what to do next. Liquid AI optimized for that full sequence.

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 Liquid AI acknowledges: larger models retain an advantage on demanding coding work. The comparisons remain vendor-run tests, so developers will need to measure tool-selection accuracy and multi-step reliability against their own workloads.

The model ships with support for llama.cpp, MLX, vLLM, SGLang and ONNX, as documented in Liquid AI's model documentation. Liquid AI's LEAP-Finetune repository covers dataset preparation, distributed training, checkpointing and exports across local GPU servers, Slurm clusters, Modal and Kubernetes environments.

Local inference changes the agent cost model

Liquid AI reports that LFM2.5-2.6B generates about 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395 while using less than 2.5 GB of memory. Liquid AI also reports roughly 30 tokens per second on a smartphone and says the model runs effectively on Raspberry Pi-class hardware.

Those measurements have not been independently verified. Exact results will depend on quantization, hardware, context length and workload. They still describe the commercial target: a model small enough to remain active without paying a cloud provider for every generated token.

That matters most for agents that monitor calendars, sort local documents, operate business software or wait for events in the background. Their economics deteriorate when every loop creates an API charge. Local execution replaces that variable inference bill with hardware and electricity already controlled by the user or enterprise.

Liquid AI is also keeping the cloud available for volume deployments. Liquid AI says the same model can approach 15,000 output tokens per second on a single Nvidia H100 under sustained concurrent load, equal to roughly 1.3 billion tokens per day. The split gives Hasani two routes to market: private execution on customer hardware and high-throughput serving where centralized infrastructure still makes sense.

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

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. Shopify's agreement with Liquid AI covers low-latency models for search and other commerce workloads.

Liquid AI has substantial capital behind the effort. It 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.

Open weights come with a revenue threshold

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 the user's legal entity reaches $10 million in annual revenue. Businesses above that threshold must purchase a commercial license from Liquid AI.

That provision gives individual developers and smaller startups wide access while preserving an enterprise licensing business for Liquid AI. It also creates an early procurement question for companies evaluating the model against alternatives distributed under Apache 2.0. Mistral's Ministral 3 3B, for example, carries Apache 2.0 terms and offers a 256K context window.

Hasani's bet is that deployment performance will carry enough value to support that licensing boundary. If LFM2.5-2.6B can reliably operate tools on low-cost CPUs, 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 weights and tooling to test that claim. The harder proof will come from agents that continue completing real work after the demo ends.