Earlier this week, the AI startup Liquid, shaped in 2023 by former MIT pc scientists, debuted LFM2.5-2.6B, a brand new open-weight language mannequin designed particularly for agentic workloads.
In launch supplies and a latest interview with VentureBeat, Liquid's researchers mentioned LFM2.5-2.6B can run completely on native {hardware} — from smartphones and laptops all the way down to a Raspberry Pi — with out counting on cloud inference or GPUs, unlocking edge AI functions and giving extra choices to enterprises working in regulated industries or with delicate info they don't wish to ship as much as the cloud.
It's finest suited to high-volume, well-defined agentic duties that run domestically — software calling, doc administration, calendar and workflow automation, and always-on background routines — and for connectivity-limited environments like autos and robotics, although coding-heavy work is healthier left to bigger fashions.
Even for these companies with out such issues, the attraction of working performant, task-specific brokers at the price of basically electrical energy, could also be sufficient to make the brand new mannequin fairly interesting.
However the customized open weights license, as with Moonshot's bigger frontier mannequin Kimi K3 launched final month, is value a detailed look by enterprise authorized groups.
The fundamentals
LFM2.5-2.6B accommodates 2.6 billion parameters, helps a 128,000-token context window, and contains native software calling. The considerably difficult title is defined by the technology of mannequin (2.5) mixed with the parameter rely (2.6B).
Each the post-trained mannequin and a base checkpoint (LFM2.5-2.6B-Base) for builders who wish to fine-tune it can be found now on Hugging Face, with day-one help for main inference stacks together with llama.cpp, MLX, vLLM, SGLang, and ONNX — positioning it for deployment throughout shopper {hardware}, enterprise infrastructure, and embedded techniques.
Liquid additionally provides an open supply fine-tuning framework, LEAP.
Quite than positioning LFM2.5-2.6B as a competitor to the biggest frontier fashions, the corporate is making a distinct argument: {that a} sufficiently succesful small mannequin can unlock classes of enterprise functions the place latency, privateness, deployment flexibility, or inference prices matter greater than absolute benchmark management.
"I do additionally imagine that the perfect fashions will likely be within the cloud, and there's no downside with that," Maxime Labonne, Liquid AI's head of post-training, instructed VentureBeat in an interview following the launch. "We wish to make fashions for an additional sort of consumer, and the easiest way of describing it’s: it’s best to use [edge AI] when you may't use a cloud mannequin."
Sufficiently small for a Raspberry Pi
Requested in regards to the minimal viable {hardware}, Labonne mentioned the mannequin runs "very, very effectively" on CPUs — and that the LFM2 structure underlying the mannequin was explicitly designed round real-world CPU efficiency quite than GPU benchmarks.
"I believe the perfect instance is a Raspberry Pi," he mentioned. "We have now plenty of demos that present that really, it really works fairly quick on the Raspberry Pi."
Firm-reported measurements point out decoding throughput of roughly 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395, whereas utilizing lower than 2.5 GB of reminiscence — and round 30 tokens per second on a smartphone. Customers can attempt the fashions on their telephones by means of Apollo, Liquid AI's cell app.
On the different finish of the deployment spectrum, Liquid AI reviews the mannequin reaches almost 15,000 output tokens per second on a single Nvidia H100 GPU underneath sustained concurrent load — roughly 1.3 billion tokens per day on one card. These figures are vendor benchmarks and haven’t been independently verified.
For Labonne, reminiscence footprint and pace usually are not conveniences however laborious constraints that decide what could be deployed in any respect.
"What we wish to present is that it's a very good trade-off, since you get the extent of high quality that you just get with a lot larger fashions, however in a tiny, tiny type issue," he mentioned. "You may deploy it in goal units the place you aren’t capable of deploy the opposite ones in any respect."
Educated for brokers as an alternative of chatbots
Liquid AI says LFM2.5-2.6B was developed across the assumption that language fashions are more and more consumed by means of agent frameworks quite than conventional conversational interfaces.
"Fashions usually are not consumed in chatbots anymore. They're actually consumed by means of agentic harnesses, like OpenClaw, like Hermes Agent," Labonne mentioned. "We needed to make it possible for this mannequin is not only good at math or at code, but it surely's good at utilizing instruments."
The mannequin is pretrained on roughly 34 trillion tokens, with a vocabulary doubled to 128K to raised help non-Latin scripts and a devoted mid-training part to increase the context window to 128K tokens for long-running agent workflows.
Submit-training follows a four-stage pipeline: supervised fine-tuning, trainer specialization (coaching separate knowledgeable fashions for domains like instruction following, math, code, and power use), multi-domain on-policy distillation (MOPD) to merge these specialists' capabilities again right into a single pupil mannequin, and eventually agentic reinforcement studying.
Throughout that final stage, the mannequin was skilled instantly inside manufacturing agent harnesses — together with Hermes Agent and OpenClaw — on practical productiveness duties involving analysis, coding, doc administration, software invocation, and workflow automation, exposing it to these harnesses' precise instruments, system prompts, and interplay patterns.
Labonne described the pipeline overhaul as producing a "completely happy accident": positive aspects that prolonged effectively past the agentic targets.
"By means of these new coaching strategies, we additionally acquired quite a bit higher at the whole lot. We acquired higher at math, at instruction following. We've by no means been good at code, really — and with this, we even acquired actually good at code," he mentioned.
Constructing the mannequin — and the harness
Notably, Liquid AI additionally constructed its personal agent harness quite than relying solely on current frameworks, and demonstrated the mannequin working inside it on a telephone, planning and calling instruments completely on-device.
"It is a harness working on a telephone, and I don't know if there's every other harness working on a telephone," Labonne mentioned.
The corporate had two causes, he defined. The primary was necessity — no phone-native harness existed. The second is a distinct interplay mannequin: immediately's harnesses anticipate a immediate, and Liquid AI desires assistants that act on their very own.
"We wish proactive brokers. We wish brokers that run within the background, test what you're doing, test your calendar, and primarily based on this context, do duties," he mentioned. "That doesn't exist immediately, actually."
Co-designing the harness and mannequin additionally lets the software program compensate for the mannequin's weak spots. "Every little thing that the mannequin is unhealthy at, the harness ought to assist the mannequin with — present as a lot help as attainable to make it extra dependable," Labonne mentioned. "Finish customers don't care if it's the mannequin or the harness. What they need is that the duty is achieved on the finish of the day."
The mannequin however works out of the field with established harnesses together with Hermes Agent, OpenClaw, and Pi, served behind any OpenAI-compatible endpoint.
Swap the harness, not the mannequin
For enterprise deployment, Labonne argued the discharge marks a shift in what small fashions can be utilized for. Till now, he mentioned, native fashions made financial sense primarily as narrowly fine-tuned specialists — skilled to do one factor at cloud-model high quality, a lot quicker and cheaper. Agentic functionality adjustments that calculus, as a result of the identical mannequin could be repurposed by altering the instruments round it quite than the mannequin itself.
"You may have a calendar assistant, and you may reuse the identical mannequin and make a gathering assistant that may file what everyone mentioned and summarize it — a bit like Granola, for instance," he mentioned. "You don't change the mannequin; you simply change the harness. You simply change the instruments round it. This offers rather more generalizability, and it's quite a bit simpler to do and quite a bit cheaper as effectively."
He nonetheless recommends fine-tuning for manufacturing deployments every time possible: "In case you don't fine-tune it, you allow some high quality on the desk. In case you fine-tune it effectively, it's going to match the efficiency of GPT and Claude — actually, in case your process will not be essentially the most advanced process on this planet," he mentioned, including that the barrier to entry has collapsed: "The bar to have the ability to do fine-tuning now’s tremendous low. It's very accessible to everybody."
The way it stacks up in opposition to DeepSeek-V4-Flash, Google's Gemma and Alibaba's Qwen
Liquid AI launched its personal benchmark comparability charts pitting LFM2.5-2.6B in opposition to the fashions enterprises are most certainly to shortlist for a 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 separate check by native AI consumer platform Atomic Chat discovered that LFM2.5-2.6B accomplished 35 software calls to finish three duties (checking climate and native time in six cities, changing one price range into six currencies, checking 4 motels and reserving for a date) 3.7 instances quicker than DeepSeek-V4-Flash (a whopping 284B parameters), the mannequin has skyrocketed to the highest of OpenRouter since its launch final week.
Gemma 4's small fashions are multimodal generalists, accepting picture and audio enter alongside textual content, and use a Per-Layer Embeddings design that retains solely a fraction of their weights lively per token — which is why Google markets them by "efficient" measurement (2.3B and 4.5B) regardless of complete footprints of 5.1B and 8B. Alibaba's Qwen3.5 small collection, launched in March, is natively multimodal from 4B up and leans on scaled reinforcement studying to chase frontier-style reasoning — Alibaba touts the 9B mannequin as matching or beating OpenAI's far bigger gpt-oss-120B on reasoning benchmarks.
LFM2.5-2.6B takes a narrower path: it’s text-only, dense, and specialised for agentic work, with Liquid AI delivery separate imaginative and prescient and audio variants of the LFM household quite than folding the whole lot into one checkpoint.
The place Qwen's post-training reinforcement studying targets reasoning, Liquid's targets software use inside actual agent harnesses.
The end result, per the corporate's revealed numbers, is that the smallest mannequin within the comparability leads each instruction-following benchmark (IFBench, Multi-IF, IFStruct) and almost each tool-use benchmark — 77.83 on ToolSandbox versus 76.44 for Qwen3.5-9B, a mannequin almost 4 instances its measurement — trailing solely that 9B mannequin on BFCLv4.
On agentic evaluations it beats each Gemma fashions throughout the board and basically ties the Qwens: 26.89 on BrowseComp+ versus 27.23 for Qwen3.5-9B. It additionally posts the perfect rating on AA Omniscience, a data benchmark that penalizes hallucination.
The Qwen fashions hold the sting the place their coaching focus lies: math (Qwen3.5-9B leads AIME25) and coding, the place bigger fashions retain a bonus on LiveCodeBench — although Labonne famous the hole is smaller than the parameter counts would recommend.
"With LiveCodeBench v6, we would not be the perfect amongst these fashions, however we're additionally by far the smallest. Displaying that we're aggressive with them is already fairly an enormous win for me," he mentioned.
One differentiator cuts the opposite means: licensing. Gemma 4 and Qwen3.5 ship underneath the permissive Apache 2.0 license — a change Google made particularly to court docket enterprises. DeepSeek-V4-Flash ships underneath a equally permissive MIT License.
In the meantime, Liquid AI's revenue-gated license (detailed under) asks bigger corporations to strike a business deal. Enterprises above the edge are successfully buying and selling license friction for footprint and tool-use efficiency.
Licensing displays a business center floor
LFM2.5-2.6B is distributed underneath the LFM Open License v1.0, which allows use, modification, and redistribution — together with business use — for organizations with lower than $10 million in annual income. Business use by bigger corporations will not be lined by the license, requiring a separate association with Liquid AI; certified nonprofits are exempt from the edge for non-commercial and analysis functions.
Labonne framed the construction as a method to maintain mannequin improvement — "the fashions are actually the moats, so we should be wise in the best way that we license them; in any other case, we can’t generate income, so we are able to't make extra fashions" — whereas characterizing the edge as a light-touch mechanism in apply.
Requested how the corporate would even know if a big enterprise quietly deployed the open weights, he was candid: "I believe this can be a query for our authorized crew, however personally, I don't know. And even if you happen to're above $10 million, the one factor that we ask you is to contact us."
The corporate pairs its licensed mannequin releases with freely revealed analysis, he added, together with new structured-output evaluations and a coaching approach that mitigates the repetition loops widespread in small fashions — a failure mode he famous Qwen fashions are "form of responsible of."
Small mannequin, massive enterprise implications
The launch coincided with an announcement from MacPaw, the Ukrainian software program firm behind CleanMyMac and Setapp, of a long-term strategic partnership with Liquid AI to construct an on-device AI stack for the Mac.
Liquid AI will design and fine-tune basis fashions for Eney, MacPaw's macOS assistant, working domestically on Apple silicon by means of MacPaw's Elix inference engine and Mnemos reminiscence layer, with outcomes anticipated later this 12 months.
Labonne pointed to the deal as a concrete validation of the scale argument: "One of many the explanation why they selected us can be as a result of the mannequin is sort of small, and so they don't have all of the reminiscence price range to run the opposite fashions."
The discharge arrives as {hardware} distributors, working system builders, and enterprise software program corporations more and more put money into native AI execution — and as agent harnesses proliferate throughout the business. Liquid AI's wager is that deployment economics, not uncooked scale, will outline an essential section of that market: brokers working repeatedly, in every single place, at zero marginal token value.
Whether or not small, extremely optimized agent fashions grow to be a major section of enterprise AI will finally rely much less on benchmark scores than on operational reliability. However Liquid AI's newest launch suggests the following aggressive frontier is not merely constructing bigger fashions — it's constructing fashions sufficiently small, and succesful sufficient, to run wherever enterprise workflows already stay.
