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 current interview with VentureBeat, Liquid’s researchers stated LFM2.5-2.6B can run solely on native {hardware} — from smartphones and laptops down to a Raspberry Pi — with out relying on cloud inference or GPUs, unlocking edge AI functions and giving extra choices to enterprises working in regulated industries or with delicate information they do not need to ship up to the cloud.
It is best fitted to high-volume, well-defined agentic duties that run regionally — device 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 higher left to bigger fashions.
Even for these companies with out such issues, the enchantment of operating performant, task-specific brokers at the value of primarily electrical energy, could also be sufficient to make the new mannequin fairly interesting.
However the custom open weights license, as with Moonshot’s larger frontier model Kimi K3 launched final month, is value an in depth look by enterprise authorized groups.
The fundamentals
LFM2.5-2.6B comprises 2.6 billion parameters, helps a 128,000-token context window, and contains native device calling. The considerably difficult title is defined by the era 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 need to fine-tune it are obtainable 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 gives an open supply fine-tuning framework, LEAP.
Somewhat than positioning LFM2.5-2.6B as a competitor to the largest frontier fashions, the firm 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 consider that the finest fashions shall be in the cloud, and there is no downside with that,” Maxime Labonne, Liquid AI’s head of post-training, advised VentureBeat in an interview following the launch. “We wish to make fashions for an additional kind of consumer, and the finest manner of describing it is: it is best to use [edge AI] when you may’t use a cloud mannequin.”
Sufficiently small for a Raspberry Pi
Requested about the minimal viable {hardware}, Labonne stated the mannequin runs “very, very nicely” on CPUs — and that the LFM2 structure underlying the mannequin was explicitly designed round real-world CPU efficiency moderately than GPU benchmarks.
“I feel the finest instance is a Raspberry Pi,” he stated. “We now have loads 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 strive the fashions on their telephones via Apollo, Liquid AI’s cellular app.
At the different finish of the deployment spectrum, Liquid AI stories the mannequin reaches practically 15,000 output tokens per second on a single Nvidia H100 GPU below sustained concurrent load — roughly 1.3 billion tokens per day on one card. These figures are vendor benchmarks and have not been independently verified.
For Labonne, reminiscence footprint and pace are not conveniences however arduous constraints that decide what could be deployed in any respect.
“What we wish to present is that it is a actually good trade-off, since you get the stage of high quality that you just get with a lot larger fashions, however in a tiny, tiny kind issue,” he stated. “You possibly can deploy it in goal units the place you are not ready to deploy the different ones in any respect.”
Skilled for brokers as a substitute of chatbots
Liquid AI says LFM2.5-2.6B was developed round the assumption that language fashions are more and more consumed via agent frameworks moderately than conventional conversational interfaces.
“Fashions are not consumed in chatbots anymore. They’re actually consumed via agentic harnesses, like OpenClaw, like Hermes Agent,” Labonne stated. “We wished to make it possible for this mannequin is not simply good at math or at code, nevertheless it’s good at utilizing instruments.”
The mannequin is pretrained on roughly 34 trillion tokens, with a vocabulary doubled to 128K to higher help non-Latin scripts and a devoted mid-training part to prolong the context window to 128K tokens for long-running agent workflows.
Put up-training follows a four-stage pipeline: supervised fine-tuning, trainer specialization (coaching separate professional fashions for domains like instruction following, math, code, and gear use), multi-domain on-policy distillation (MOPD) to merge these consultants’ capabilities again right into a single scholar mannequin, and at last agentic reinforcement studying.
Throughout that final stage, the mannequin was educated instantly inside manufacturing agent harnesses — together with Hermes Agent and OpenClaw — on life like productiveness duties involving analysis, coding, doc administration, device invocation, and workflow automation, exposing it to these harnesses’ precise instruments, system prompts, and interplay patterns.
Labonne described the pipeline overhaul as producing a “blissful accident”: positive aspects that prolonged nicely past the agentic targets.
“By means of these new coaching methods, we additionally received so much higher at every thing. We received higher at math, at instruction following. We have by no means been good at code, truly — and with this, we even received actually good at code,” he stated.
Constructing the mannequin — and the harness
Notably, Liquid AI additionally constructed its personal agent harness moderately than relying solely on current frameworks, and demonstrated the mannequin operating inside it on a cellphone, planning and calling instruments solely on-device.
“This is a harness operating on a cellphone, and I do not know if there’s another harness operating on a cellphone,” Labonne stated.
The corporate had two causes, he defined. The primary was necessity — no phone-native harness existed. The second is a distinct interplay mannequin: as we speak’s harnesses look forward to a immediate, and Liquid AI desires assistants that act on their very own.
“We wish proactive brokers. We wish brokers that run in the background, verify what you are doing, verify your calendar, and based mostly on this context, do duties,” he stated. “That does not exist as we speak, actually.”
Co-designing the harness and mannequin additionally lets the software program compensate for the mannequin’s weak spots. “The whole lot that the mannequin is unhealthy at, the harness ought to assist the mannequin with — present as a lot help as potential to make it extra dependable,” Labonne stated. “Finish customers do not care if it is the mannequin or the harness. What they need is that the activity is achieved at 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 launch marks a shift in what small fashions can be utilized for. Till now, he stated, native fashions made financial sense primarily as narrowly fine-tuned specialists — educated 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 moderately than the mannequin itself.
“You possibly can have a calendar assistant, and you may reuse the identical mannequin and make a gathering assistant that can file what everyone stated and summarize it — a bit like Granola, for instance,” he stated. “You do not change the mannequin; you simply change the harness. You simply change the instruments round it. This offers rather more generalizability, and it is so much simpler to do and so much cheaper as nicely.”
He nonetheless recommends fine-tuning for manufacturing deployments at any time when possible: “In case you do not fine-tune it, you permit some high quality on the desk. In case you fine-tune it nicely, it is going to match the efficiency of GPT and Claude — actually, in case your activity is not the most advanced activity in the world,” he stated, including that the barrier to entry has collapsed: “The bar to give you the option to do fine-tuning now is tremendous low. It is very accessible to everybody.”
The way it stacks up towards DeepSeek-V4-Flash, Google’s Gemma and Alibaba’s Qwen
Liquid AI launched its personal benchmark comparability charts pitting LFM2.5-2.6B towards the fashions enterprises are probably to shortlist for the identical 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 take a look at by native AI shopper platform Atomic Chat discovered that LFM2.5-2.6B accomplished 35 device calls to full three duties (checking climate and native time in six cities, changing one funds 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 high 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” dimension (2.3B and 4.5B) regardless of whole footprints of 5.1B and 8B. Alibaba’s Qwen3.5 small collection, released 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 is text-only, dense, and specialised for agentic work, with Liquid AI delivery separate imaginative and prescient and audio variants of the LFM household moderately than folding every thing into one checkpoint.
The place Qwen’s post-training reinforcement studying targets reasoning, Liquid’s targets device use inside actual agent harnesses.
The outcome, per the firm’s printed numbers, is that the smallest mannequin in the comparability leads each instruction-following benchmark (IFBench, Multi-IF, IFStruct) and practically each tool-use benchmark — 77.83 on ToolSandbox versus 76.44 for Qwen3.5-9B, a mannequin practically 4 instances its dimension — trailing solely that 9B mannequin on BFCLv4.
On agentic evaluations it beats each Gemma fashions throughout the board and primarily ties the Qwens: 26.89 on BrowseComp+ versus 27.23 for Qwen3.5-9B. It additionally posts the finest rating on AA Omniscience, a data benchmark that penalizes hallucination.
The Qwen fashions maintain the edge 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 counsel.
“With LiveCodeBench v6, we’d not be the finest amongst these fashions, however we’re additionally by far the smallest. Displaying that we’re aggressive with them is already fairly a giant win for me,” he stated.
One differentiator cuts the different manner: licensing. Gemma 4 and Qwen3.5 ship below the permissive Apache 2.0 license — a change Google made specifically to court enterprises. DeepSeek-V4-Flash ships under a similarly permissive MIT License.
In the meantime, Liquid AI’s revenue-gated license (detailed under) asks bigger firms to strike a industrial deal. Enterprises above the threshold are successfully buying and selling license friction for footprint and tool-use efficiency.
Licensing displays a industrial center floor
LFM2.5-2.6B is distributed below the LFM Open License v1.0, which allows use, modification, and redistribution — together with industrial use — for organizations with lower than $10 million in annual income. Industrial use by bigger firms is not coated by the license, requiring a separate association with Liquid AI; certified nonprofits are exempt from the threshold for non-commercial and analysis functions.
Labonne framed the construction as a manner to maintain mannequin growth — “the fashions are actually the moats, so we want to be smart in the manner that we license them; in any other case, we can not earn cash, so we won’t make extra fashions” — whereas characterizing the threshold as a light-touch mechanism in observe.
Requested how the firm would even know if a big enterprise quietly deployed the open weights, he was candid: “I feel this is a query for our authorized staff, however personally, I do not know. And even when you’re above $10 million, the solely factor that we ask you is to contact us.”
The corporate pairs its licensed mannequin releases with freely printed analysis, he added, together with new structured-output evaluations and a coaching approach that mitigates the repetition loops frequent 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, operating regionally on Apple silicon via 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 dimension argument: “Considered one of the the explanation why they selected us is additionally as a result of the mannequin is fairly small, and so they haven’t got all the reminiscence funds to run the different fashions.”
The discharge arrives as {hardware} distributors, working system builders, and enterprise software program firms more and more spend money on native AI execution — and as agent harnesses proliferate throughout the trade. Liquid AI’s guess is that deployment economics, not uncooked scale, will outline an vital section of that market: brokers operating repeatedly, in every single place, at zero marginal token value.
Whether or not small, extremely optimized agent fashions develop into a big section of enterprise AI will in the end rely much less on benchmark scores than on operational reliability. However Liquid AI’s newest launch suggests the subsequent aggressive frontier is now not merely constructing bigger fashions — it is constructing fashions sufficiently small, and succesful sufficient, to run wherever enterprise workflows already reside.
Disclaimer: This article is sourced from external platforms. OverBeta has not independently verified the information. Readers are advised to verify details before relying on them.
