Liquid AI's LFM2.5-2.6B: Small Models, Big Edge Promises
Liquid AI, the startup spun out of MIT in 2023 by a team of former computer scientists, just shipped something that might quietly reshape how enterprises think about deploying AI at the edge. LFM2.5-2.6B is a 2.6 billion parameter open-weight language model that runs entirely on local hardware — phones, laptops, even a Raspberry Pi — without needing cloud inference or a GPU. The name's a mouthful (generation 2.5 meets 2.6 billion parameters), but the pitch is straightforward: build models for people who can't use cloud models.
The model supports a 128,000-token context window, includes native tool calling, and ships with both a post-trained version and a base checkpoint (LFM2.5-2.6B-Base) for developers who want to fine-tune it themselves. Day-one inference support covers llama.cpp, MLX, vLLM, SGLang, and ONNX — covering most of the major stacks you'd actually use on consumer or embedded hardware. Liquid also released LEAP, an open-source fine-tuning framework, alongside it.
Why Edge AI Matters Now
The obvious use case here isn't replacing GPT-level models. It's the stuff you can't send to the cloud — regulated industries, healthcare, defense, or any org that simply doesn't want sensitive data leaving their premises. Liquid's head of post-training, Maxime Labonne, put it bluntly in a VentureBeat interview: "You should use edge AI when you can't use a cloud model."
That's not a niche concern anymore. Hardware vendors, OS developers, and enterprise software companies are all investing heavily in local AI execution. Agent harnesses are proliferating across the industry. Liquid's bet is that deployment economics — latency, privacy, inference cost — will define an increasingly important segment of that market.
The model targets high-volume, well-defined agentic tasks: tool calling, document management, calendar and workflow automation, always-on background routines, and connectivity-limited environments like vehicles and robotics. Coding-heavy work? Labonne admits larger models still win there. But for most enterprise productivity tasks, the trade-off looks compelling.
Hardware Numbers That Actually Impress
Liquid's own benchmarks tell a story worth paying attention to. The model decodes at roughly 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395, using less than 2.5 GB of memory. On a smartphone, it hits about 30 tokens per second — slow by cloud standards, but functional for background agent work. Users can test it themselves through Apollo, Liquid's mobile app.
At the other end of the spectrum, Liquid reports nearly 15,000 output tokens per second on a single Nvidia H100 GPU under sustained concurrent load, roughly 1.3 billion tokens per day on one card. These are vendor benchmarks, not independently verified, but they're still useful for understanding the model's range.
What's genuinely interesting is the Raspberry Pi performance. Labonne said it runs "very, very well" on CPUs, and the LFM2 architecture was explicitly designed around real-world CPU performance rather than GPU benchmarks. That's a departure from how most small model releases position themselves, they chase benchmark scores, not whether something actually works on hardware you can buy at a store.
Training for Agents, Not Chatbots
Liquid's approach to training LFM2.5-2.6B reflects a shifting assumption about how language models get used. "Models are not consumed in chatbots anymore," Labonne told VentureBeat. "They're really consumed through agentic harnesses, like OpenClaw, like Hermes Agent."
The model was pretrained on approximately 34 trillion tokens with a vocabulary doubled to 128K to better support non-Latin scripts and long-running agent workflows. Post-training follows a four-stage pipeline: supervised fine-tuning, teacher specialization (separate expert models for domains like instruction following, math, code, and tool use), multi-domain on-policy distillation (MOPD) to merge those experts back into a single student model, and finally agentic reinforcement learning.
Here's where it gets interesting: during that last stage, the model was trained directly inside production agent harnesses, Hermes Agent and OpenClaw, on realistic productivity tasks involving research, coding, document management, tool invocation, and workflow automation. It was exposed to those harnesses' actual tools, system prompts, and interaction patterns.
Labonne called the pipeline overhaul a "happy accident", gains that extended well beyond the agentic targets. "We got better at math, at instruction following," he said. "We've never been good at code, actually, and with this, we even got really good at code."
The Harness Strategy
Liquid didn't just build a model. It also built its own agent harness, demonstrating it running on a phone and planning tool calls entirely on-device. "I don't know if there's any other harness running on a phone," Labonne said.
Two reasons drove this decision. First, necessity, no phone-native harness existed. Second, a different interaction model: current harnesses wait for a prompt, and Liquid wants proactive agents that run in the background, check your calendar, and act on context. "We want proactive agents," Labonne said. "That doesn't exist today, really."
Co-designing harness and model lets the software compensate for the model's weak spots. "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," Labonne explained. "End users don't care if it's the model or the harness. What they want is that the task is achieved ultimately."
The model works with established harnesses too, Hermes Agent, OpenClaw, and Pi, served behind any OpenAI-compatible endpoint.
Swap the Harness, Not the Model
For enterprise deployment, Labonne argued this shifts what small models can be used for. Historically, local models made economic sense mainly as narrowly fine-tuned specialists, trained to do one thing at cloud-model quality, faster and cheaper. Agentic capability changes that calculus because the same model can be repurposed by swapping tools around rather than changing the model itself.
"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 said. "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."
He still recommends fine-tuning for production deployments when 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 said, adding that the barrier to entry has collapsed: "The bar to be able to do fine-tuning now is super low. It's very accessible to everyone."
How It Stacks Up Against the Competition
Liquid released benchmark comparisons pitting LFM2.5-2.6B against models enterprises are likely to shortlist for 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 separate test by local AI client platform Atomic Chat found LFM2.5-2.6B completed 35 tool calls across three tasks (checking weather and local time in six cities, converting one budget into six currencies, checking four hotels and booking for a date) 3.7 times faster than DeepSeek-V4-Flash, a 284B parameter model. Since that release, the model has skyrocketed to the top of OpenRouter.
Gemma 4's small models are multimodal generalists using Per-Layer Embeddings to keep only a fraction of weights active per token. Alibaba's Qwen3.5 series is natively multimodal from 4B up and leans on scaled reinforcement learning for frontier-style reasoning.
LFM2.5-2.6B takes a narrower path: text-only, dense, and specialized for agentic work. Liquid ships separate vision and audio variants of the LFM family rather than folding everything into one checkpoint.
Per Liquid's published numbers, the smallest model in the comparison leads every instruction-following benchmark (IFBench, Multi-IF, IFStruct) and nearly every tool-use benchmark, 77.83 on ToolSandbox versus 76.44 for Qwen3.5-9B, a model nearly four times its size. It trails only Qwen3.5-9B on BFCLv4. On agentic evaluations, it beats both Gemma models across the board and essentially ties the Qwens: 26.89 on BrowseComp+ versus 27.23 for Qwen3.5-9B. It also posts the best score on AA Omniscience, a knowledge benchmark that penalizes hallucination.
The Qwen models keep the edge where their training focus lies: math (Qwen3.5-9B leads AIME25) and coding, where larger models retain an advantage on LiveCodeBench. "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," Labonne said.
The Licensing Question
Here's where things get tricky for enterprise legal teams. Gemma 4 and Qwen3.5 ship under Apache 2.0, Google made that change specifically to court enterprises. DeepSeek-V4-Flash ships under MIT. LFM2.5-2.6B uses the LFM Open License v1.0, which permits use, modification, and redistribution, including commercial use, for organizations with less than $10 million in annual revenue.
Commercial use by larger companies requires a separate arrangement with Liquid AI. Qualified nonprofits are exempt from the threshold for non-commercial and research purposes.
Labonne framed the structure as necessary for sustainability: "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." When asked how the company would even know if a large enterprise quietly deployed the open weights, he 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."
It's a revenue-gated license that asks larger companies to strike a commercial deal. Enterprises above the threshold are effectively trading license friction for footprint and tool-use performance. Liquid pairs its licensed model releases with freely published research, including new structured-output evaluations and a training technique that mitigates the repetition loops common in small models — a failure mode Labonne noted Qwen models are "kind of guilty of."
The MacPaw Partnership
The launch coincided with an announcement from MacPaw, the Ukrainian software company behind CleanMyMac and Setapp, of a long-term strategic partnership with Liquid AI to build an on-device AI stack for the Mac.
Liquid AI will design and fine-tune foundation models for Eney, MacPaw's macOS assistant, running locally on Apple silicon through MacPaw's Elix inference engine and Mnemos memory layer, with results expected later this year.
Labonne pointed to the deal as a concrete validation of the size 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."
What This Means for Enterprise AI
The release arrives as hardware vendors, operating system developers, and enterprise software companies increasingly invest in local AI execution — and as agent harnesses proliferate across the industry. Liquid AI's bet is that deployment economics, not raw scale, will define an important segment of that market: agents running continuously, everywhere, at zero marginal token cost.
Whether small, highly optimized agent models become a significant segment of enterprise AI will ultimately depend less on benchmark scores than on operational reliability. But Liquid AI's latest release suggests the next competitive frontier is no longer simply building larger models — it's building models small enough, and capable enough, to run wherever enterprise workflows already live.
For enterprises sitting on the fence, the message is clear: edge AI is no longer a compromise. It's a deliberate architectural choice, and LFM2.5-2.6B makes a compelling case for it.