By admin 
Liquid AI, an ambitious AI startup founded in 2023 by a cadre of distinguished former MIT computer scientists, has unveiled LFM2.5-2.6B, a novel open-weight language model specifically engineered to excel in agentic workloads. This release marks a significant stride in democratizing advanced AI capabilities, pushing the boundaries of what can be achieved with localized, efficient computation. The model’s architecture and performance characteristics suggest a paradigm shift, moving away from the monolithic cloud-based AI solutions towards a future of pervasive, on-device intelligence.
In their official release materials and a recent in-depth interview with VentureBeat, Liquid AI’s research team articulated the core philosophy behind LFM2.5-2.6B: unparalleled accessibility and operational flexibility. A key differentiator of this new model is its ability to run entirely on local hardware, a capability that spans the spectrum from high-end smartphones and laptops to surprisingly modest devices like the Raspberry Pi. This eliminates the dependency on costly cloud inference infrastructure and power-hungry GPUs, thereby unlocking a vast array of edge AI applications. For enterprises operating within highly regulated industries or handling sensitive information, the ability to process data locally, without the need for transmission to external servers, offers a compelling solution for enhanced privacy and security.
LFM2.5-2.6B is particularly well-suited for a defined set of agentic tasks that benefit from local execution. These include high-volume, well-defined operations such as intricate tool calling, efficient document management, seamless calendar and workflow automation, and the implementation of always-on background routines. Its efficiency also makes it an ideal candidate for environments where connectivity is a persistent challenge, such as in autonomous vehicles and complex robotics systems. While not designed to compete with the largest frontier models for highly coding-intensive or complex reasoning tasks, its specialization in agentic functions presents a distinct advantage. Even for businesses without the stringent privacy concerns of regulated industries, the prospect of deploying performant, task-specific agents that operate at the marginal cost of electricity is a powerful economic incentive, making LFM2.5-2.6B an exceptionally attractive proposition.
However, as with any release of open-weight models, particularly those with custom licensing agreements, a thorough review by enterprise legal teams is strongly advised. This is particularly true given the precedent set by Moonshot’s recent release of its larger frontier model, Kimi K3, which also featured an open-weight model with a nuanced license. Understanding the specific terms and conditions of Liquid AI’s custom open weights license is paramount to ensure compliant and effective deployment within an enterprise context.
At its core, LFM2.5-2.6B boasts 2.6 billion parameters, a substantial 128,000-token context window, and native support for tool calling. The model’s designation, LFM2.5-2.6B, elegantly reflects its lineage (2.5 generation) and its parameter count (2.6 billion). Liquid AI has made both the post-trained model, ready for immediate deployment, and a base checkpoint (LFM2.5-2.6B-Base) available for developers keen on fine-tuning the model for their specific needs. Both are accessible on Hugging Face, a leading platform for AI model sharing. Crucially, the model offers day-one support for major inference stacks, including llama.cpp, MLX, vLLM, SGLang, and ONNX. This broad compatibility ensures seamless integration across a diverse range of deployment environments, from consumer-grade hardware and robust enterprise infrastructure to highly specialized embedded systems. To further empower developers, Liquid AI has also released LEAP, an open-source framework designed for efficient model fine-tuning.
Liquid AI strategically positions LFM2.5-2.6B not as a direct competitor to the behemoth frontier models, but as an enabler for a new class of enterprise applications. The company’s argument hinges on the notion that a sufficiently capable, yet compact, model can unlock solutions where latency, data privacy, deployment flexibility, and inference costs are more critical than achieving absolute leadership on broad benchmark tests. Maxime Labonne, Liquid AI’s Head of Post-Training, emphasized this point in his interview with VentureBeat: "I do also believe that the best models will be in the cloud, and there’s no problem with that. We want to make models for another type of user, and the best way of describing it is: you should use [edge AI] when you can’t use a cloud model." This highlights a clear strategy to address the limitations of cloud-centric AI and cater to a distinct set of user needs.
The claim that LFM2.5-2.6B can run on hardware as modest as a Raspberry Pi is not an exaggeration. When questioned about the minimum viable hardware, Labonne confirmed that the model performs "very, very well" on CPUs. He further elaborated that the underlying LFM2 architecture was meticulously designed with real-world CPU performance as a primary consideration, rather than solely focusing on GPU benchmarks. "I think the best example is a Raspberry Pi," he stated. "We have a lot of demos that show that actually, it works pretty fast on the Raspberry Pi."
Quantifiable performance metrics underscore these claims. Liquid AI reports decoding throughputs of approximately 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395, all while consuming less than 2.5 GB of memory. Even on a standard smartphone, the model achieves a respectable 30 tokens per second. Users can experience this on-device capability firsthand through Apollo, Liquid AI’s mobile application. For scenarios requiring higher throughput, the model demonstrates impressive scalability. On a single Nvidia H100 GPU, Liquid AI reports that LFM2.5-2.6B can achieve nearly 15,000 output tokens per second under sustained concurrent load, translating to roughly 1.3 billion tokens daily on a single card. While these figures are vendor benchmarks and await independent verification, they paint a picture of remarkable efficiency and adaptability. For Labonne, the low memory footprint and high speed are not mere conveniences but fundamental constraints that dictate the feasibility of deploying AI in diverse environments. "What we want to show is that it’s a really good trade-off, because you get the level of quality that you get with much bigger models, but in a tiny, tiny form factor," he explained. "You can deploy it in target devices where you are not able to deploy the other ones at all."
A significant aspect of LFM2.5-2.6B’s design philosophy is its specialization for agentic tasks rather than traditional chatbot interfaces. Liquid AI operates under the premise that the future of language model consumption lies in agent frameworks. "Models are not consumed in chatbots anymore. They’re really consumed through agentic harnesses, like OpenClaw, like Hermes Agent," Labonne noted. "We wanted to make sure that this model is not just good at math or at code, but it’s good at using tools." To achieve this, the model was pre-trained on an immense dataset of approximately 34 trillion tokens. Its vocabulary was intentionally doubled to 128,000 tokens to enhance support for non-Latin scripts. Furthermore, a dedicated mid-training phase was implemented to extend the context window to a significant 128,000 tokens, facilitating the execution of long-running agent workflows.
The post-training pipeline for LFM2.5-2.6B is a sophisticated, four-stage process. It begins with supervised fine-tuning, followed by teacher specialization, where separate expert models are trained for specific domains like instruction following, mathematics, coding, and tool usage. This is succeeded by multi-domain on-policy distillation (MOPD), a technique used to merge the capabilities of these specialized expert models into a single, unified student model. The final stage involves agentic reinforcement learning. During this crucial phase, the model was trained directly within production agent harnesses, including Hermes Agent and OpenClaw. This hands-on training exposed the model to realistic productivity tasks encompassing research, coding, document management, tool invocation, and workflow automation, allowing it to learn from the harnesses’ actual tools, system prompts, and interaction patterns. Labonne described this overhaul of the training pipeline as a "happy accident," yielding unexpected improvements that extended beyond the initial agentic targets. "Through these new training techniques, we also got a lot better at everything. We got better at math, at instruction following. We’ve never been good at code, actually – and with this, we even got really good at code," he revealed.
Liquid AI’s commitment to the agentic paradigm extends to the development of its own agent harness, a strategic move to complement and optimize the performance of LFM2.5-2.6B. Rather than relying exclusively on existing frameworks, the company built its own, demonstrating the model running natively on a phone, capable of planning and invoking tools entirely on-device. "This is a harness running on a phone, and I don’t know if there’s any other harness running on a phone," Labonne remarked. This dual development approach was driven by two primary factors. Firstly, the absence of a phone-native harness necessitated its creation. Secondly, Liquid AI envisions a different interaction model for AI assistants – one where they are proactive rather than solely reactive to prompts. "We want proactive agents. We want agents that run in the background, check what you’re doing, check your calendar, and based on this context, do tasks," Labonne explained. "That doesn’t exist today, really." The co-design of the harness and the model allows for a synergistic relationship where the software can actively compensate for the model’s inherent limitations. "Everything that the model is bad at, the harness should help the model with – provide as much assistance as possible to make it more reliable," he stated. "End users don’t care if it’s the model or the harness. What they want is that the task is achieved at the end of the day." Despite developing its own harness, LFM2.5-2.6B remains compatible with established frameworks such as Hermes Agent, OpenClaw, and Pi, and can be served through any OpenAI-compatible endpoint.
The "swap the harness, not the model" philosophy is a cornerstone of Liquid AI’s enterprise deployment strategy. Labonne argues that this release signifies a fundamental shift in the utility of small models. Previously, local models were primarily economically viable as narrowly fine-tuned specialists, excelling at a single task with cloud-model quality but at a significantly reduced cost and latency. The introduction of agentic capabilities changes this calculus, enabling the repurposing of a single model for diverse applications simply by altering the surrounding tools and harness. "You can have a calendar assistant, and you can reuse the same model and make a meeting assistant that will record what everybody said and summarize it – a bit like Granola, for example," Labonne illustrated. "You don’t change the model; you just change the harness. You just change the tools around it. This gives much more generalizability, and it’s a lot easier to do and a lot cheaper as well." While acknowledging the advantages of this flexibility, Labonne still strongly recommends fine-tuning for production deployments whenever feasible. "If you don’t fine-tune it, you leave some quality on the table. If you fine-tune it well, it’s going to match the performance of GPT and Claude – really, if your task is not the most complex task in the world," he asserted, noting that the barrier to entry for effective fine-tuning has been dramatically lowered, making it "super low" and "very accessible to everyone."
In its release materials, Liquid AI provided a detailed benchmark comparison, positioning LFM2.5-2.6B against prominent models favored for similar edge deployments: Google’s Gemma 4 E2B (5.1B parameters) and E4B (8B), and Alibaba’s Qwen3.5-4B (4.7B) and Qwen3.5-9B (9.7B). A compelling independent evaluation by Atomic Chat, a local AI client platform, revealed that LFM2.5-2.6B completed a complex sequence of tasks—checking weather and local time in six cities, converting a budget into six currencies, and checking four hotels for booking availability—3.7 times faster than DeepSeek-V4-Flash, a model with a staggering 284 billion parameters. This performance has propelled LFM2.5-2.6B to the top of OpenRouter rankings since its release.
The comparison with Gemma and Qwen models highlights LFM2.5-2.6B’s specialized strengths. Google’s Gemma 4 models are multimodal generalists, capable of processing image and audio inputs alongside text. They employ a Per-Layer Embeddings design, activating only a fraction of their weights per token, which allows Google to market them by "effective" size (2.3B and 4.5B) despite their larger total footprints. Alibaba’s Qwen3.5 small series, launched in March, are natively multimodal from the 4B parameter version upwards and leverage scaled reinforcement learning to emulate frontier-level reasoning. Alibaba claims their 9B model matches or surpasses OpenAI’s significantly larger gpt-oss-120B on reasoning benchmarks.
LFM2.5-2.6B, in contrast, pursues a more focused strategy: it is text-only, dense, and deliberately specialized for agentic workflows. Liquid AI plans to release separate vision and audio variants of the LFM family, rather than integrating these capabilities into a single checkpoint. While Qwen’s post-training reinforcement learning targets general reasoning, Liquid AI’s methodology is specifically geared towards enhancing tool use within actual agent harnesses. The results, as presented by Liquid AI, are striking: the smallest model in the comparison leads on all instruction-following benchmarks (IFBench, Multi-IF, IFStruct) and on nearly all tool-use benchmarks. It achieved a score of 77.83 on ToolSandbox, surpassing Qwen3.5-9B’s 76.44, despite the Qwen model being nearly four times its size. LFM2.5-2.6B trailed only the 9B Qwen model on BFCLv4.
However, the comparison reveals areas where larger, more generalist models retain an advantage. The Qwen models, with their focus on reasoning and broader training, lead in specific domains like mathematics (Qwen3.5-9B excels on AIME25) and coding. On LiveCodeBench, larger models still demonstrate superior performance. Labonne, however, points out that the performance gap in coding is considerably smaller than the disparity in parameter counts would suggest. "With LiveCodeBench v6, we might not be the best among these models, but we’re also by far the smallest. Showing that we’re competitive with them is already quite a big win for me," he commented.
A critical differentiator in the competitive landscape is licensing. Gemma 4 and Qwen3.5 are distributed under the permissive Apache 2.0 license, a move Google explicitly made to attract enterprise adoption. DeepSeek-V4-Flash also operates under a similarly permissive MIT License. In contrast, Liquid AI employs a revenue-gated license, requiring larger companies to establish a commercial agreement. This licensing structure positions enterprises as trading potential license friction for the model’s distinct advantages in footprint and tool-use performance.
Liquid AI’s licensing strategy, embodied in the LFM Open License v1.0, strikes a commercial middle ground. The license permits use, modification, and redistribution, including for commercial purposes, for organizations with annual revenues under $10 million. Commercial use by companies exceeding this threshold necessitates a separate arrangement with Liquid AI. Qualified non-profits are exempt from this revenue threshold for non-commercial and research activities. Labonne frames this structure as essential for sustaining ongoing model development. "The models are really the moats, so we need to be sensible in the way that we license them; otherwise, we cannot make money, so we can’t make more models," he explained. He characterizes the revenue threshold as a "light-touch mechanism in practice." When queried about enforcement for large enterprises quietly deploying the open weights, Labonne was candid: "I think this is a question for our legal team, but personally, I don’t know. And even if you’re above $10 million, the only thing that we ask you is to contact us." The company complements its licensed model releases with openly published research, including novel structured-output evaluations and a training technique designed to mitigate the repetition loops that often plague smaller models—a failure mode he notes that Qwen models are "kind of guilty of."
The significance of Liquid AI’s release is amplified by concurrent developments. The launch coincided with an announcement detailing a long-term strategic partnership between Liquid AI and MacPaw, the Ukrainian software company renowned for products like CleanMyMac and Setapp. This collaboration aims to develop an on-device AI stack specifically for macOS. Liquid AI will be responsible for designing and fine-tuning foundation models for Eney, MacPaw’s macOS assistant. These models will operate locally on Apple silicon, leveraging MacPaw’s Elix inference engine and Mnemos memory layer, with initial results anticipated later this year. Labonne highlighted this partnership as a tangible validation of the size-first argument: "One of the reasons why they chose us is also because the model is quite small, and they don’t have all the memory budget to run the other models."
The release of LFM2.5-2.6B arrives at a pivotal moment, as hardware vendors, operating system developers, and enterprise software companies increasingly prioritize and invest in local AI execution. Simultaneously, the proliferation of agent harnesses across the industry signals a growing demand for more sophisticated AI agents. Liquid AI’s strategic bet is that deployment economics, rather than sheer model scale, will ultimately define a crucial segment of this burgeoning market: agents that can run continuously, ubiquitously, and at a zero marginal token cost. The ultimate success of small, highly optimized agent models in enterprise settings will likely hinge less on benchmark scores and more on operational reliability and seamless integration. However, Liquid AI’s latest offering strongly suggests that the next competitive frontier in artificial intelligence is no longer solely about constructing larger models, but about engineering models that are sufficiently compact and capable to operate natively within the existing workflows and environments where enterprise tasks already reside.
