When Liquid AI, a startup founded by MIT pc scientists again in 2023, launched its Liquid Basis Fashions collection 2 (LFM2) in July 2025, the pitch was simple: ship the quickest on-device basis fashions available on the market utilizing the brand new "liquid" structure, with coaching and inference effectivity that made small fashions a critical various to cloud-only giant language fashions (LLMs) comparable to OpenAI's GPT collection and Google's Gemini.
The preliminary launch shipped dense checkpoints at 350M, 700M, and 1.2B parameters, a hybrid structure closely weighted towards gated quick convolutions, and benchmark numbers that positioned LFM2 forward of equally sized rivals like Qwen3, Llama 3.2, and Gemma 3 on each high quality and CPU throughput. The message to enterprises was clear: real-time, privacy-preserving AI on telephones, laptops, and autos now not required sacrificing functionality for latency.
Within the months since that launch, Liquid has expanded LFM2 right into a broader product line — including task-and-domain-specialized variants, a small video ingestion and evaluation mannequin, and an edge-focused deployment stack known as LEAP — and positioned the fashions because the management layer for on-device and on-prem agentic programs.
Now, with the publication of the detailed, 51-page LFM2 technical report on arXiv, the corporate goes a step additional: making public the structure search course of, coaching information combination, distillation goal, curriculum technique, and post-training pipeline behind these fashions.
And in contrast to earlier open fashions, LFM2 is constructed round a repeatable recipe: a hardware-in-the-loop search course of, a coaching curriculum that compensates for smaller parameter budgets, and a post-training pipeline tuned for instruction following and power use.
Fairly than simply providing weights and an API, Liquid is successfully publishing an in depth blueprint that different organizations can use as a reference for coaching their very own small, environment friendly fashions from scratch, tuned to their very own {hardware} and deployment constraints.
A mannequin household designed round actual constraints, not GPU labs
The technical report begins with a premise enterprises are intimately aware of: actual AI programs hit limits lengthy earlier than benchmarks do. Latency budgets, peak reminiscence ceilings, and thermal throttling outline what can truly run in manufacturing—particularly on laptops, tablets, commodity servers, and cellular gadgets.
To handle this, Liquid AI carried out structure search immediately on course {hardware}, together with Snapdragon cellular SoCs and Ryzen laptop computer CPUs. The result’s a constant end result throughout sizes: a minimal hybrid structure dominated by gated quick convolution blocks and a small variety of grouped-query consideration (GQA) layers. This design was repeatedly chosen over extra unique linear-attention and SSM hybrids as a result of it delivered a greater quality-latency-memory Pareto profile underneath actual system circumstances.
This issues for enterprise groups in 3 ways:
-
Predictability. The structure is straightforward, parameter-efficient, and steady throughout mannequin sizes from 350M to 2.6B.
-
Operational portability. Dense and MoE variants share the identical structural spine, simplifying deployment throughout combined {hardware} fleets.
-
On-device feasibility. Prefill and decode throughput on CPUs surpass comparable open fashions by roughly 2× in lots of circumstances, lowering the necessity to offload routine duties to cloud inference endpoints.
As an alternative of optimizing for educational novelty, the report reads as a scientific try and design fashions enterprises can truly ship.
That is notable and extra sensible for enterprises in a discipline the place many open fashions quietly assume entry to multi-H100 clusters throughout inference.
A coaching pipeline tuned for enterprise-relevant conduct
LFM2 adopts a coaching strategy that compensates for the smaller scale of its fashions with construction slightly than brute drive. Key parts embrace:
-
10–12T token pre-training and an extra 32K-context mid-training section, which extends the mannequin’s helpful context window with out exploding compute prices.
-
A decoupled High-Okay information distillation goal that sidesteps the instability of ordinary KL distillation when lecturers present solely partial logits.
-
A three-stage post-training sequence—SFT, length-normalized choice alignment, and mannequin merging—designed to provide extra dependable instruction following and tool-use conduct.
For enterprise AI builders, the importance is that LFM2 fashions behave much less like “tiny LLMs” and extra like sensible brokers in a position to comply with structured codecs, adhere to JSON schemas, and handle multi-turn chat flows. Many open fashions at comparable sizes fail not as a consequence of lack of reasoning potential, however as a consequence of brittle adherence to instruction templates. The LFM2 post-training recipe immediately targets these tough edges.
In different phrases: Liquid AI optimized small fashions for operational reliability, not simply scoreboards.
Multimodality designed for system constraints, not lab demos
The LFM2-VL and LFM2-Audio variants replicate one other shift: multimodality constructed round token effectivity.
Fairly than embedding an enormous imaginative and prescient transformer immediately into an LLM, LFM2-VL attaches a SigLIP2 encoder via a connector that aggressively reduces visible token rely through PixelUnshuffle. Excessive-resolution inputs robotically set off dynamic tiling, preserving token budgets controllable even on cellular {hardware}. LFM2-Audio makes use of a bifurcated audio path—one for embeddings, one for technology—supporting real-time transcription or speech-to-speech on modest CPUs.
For enterprise platform architects, this design factors towards a sensible future the place:
-
doc understanding occurs immediately on endpoints comparable to discipline gadgets;
-
audio transcription and speech brokers run regionally for privateness compliance;
-
multimodal brokers function inside mounted latency envelopes with out streaming information off-device.
The through-line is identical: multimodal functionality with out requiring a GPU farm.
Retrieval fashions constructed for agent programs, not legacy search
LFM2-ColBERT extends late-interaction retrieval right into a footprint sufficiently small for enterprise deployments that want multilingual RAG with out the overhead of specialised vector DB accelerators.
That is notably significant as organizations start to orchestrate fleets of brokers. Quick native retrieval—working on the identical {hardware} because the reasoning mannequin—reduces latency and supplies a governance win: paperwork by no means go away the system boundary.
Taken collectively, the VL, Audio, and ColBERT variants present LFM2 as a modular system, not a single mannequin drop.
The rising blueprint for hybrid enterprise AI architectures
Throughout all variants, the LFM2 report implicitly sketches what tomorrow’s enterprise AI stack will appear to be: hybrid local-cloud orchestration, the place small, quick fashions working on gadgets deal with time-critical notion, formatting, software invocation, and judgment duties, whereas bigger fashions within the cloud provide heavyweight reasoning when wanted.
A number of developments converge right here:
-
Value management. Operating routine inference regionally avoids unpredictable cloud billing.
-
Latency determinism. TTFT and decode stability matter in agent workflows; on-device eliminates community jitter.
-
Governance and compliance. Native execution simplifies PII dealing with, information residency, and auditability.
-
Resilience. Agentic programs degrade gracefully if the cloud path turns into unavailable.
Enterprises adopting these architectures will probably deal with small on-device fashions because the “management airplane” of agentic workflows, with giant cloud fashions serving as on-demand accelerators.
LFM2 is likely one of the clearest open-source foundations for that management layer so far.
The strategic takeaway: on-device AI is now a design alternative, not a compromise
For years, organizations constructing AI options have accepted that “actual AI” requires cloud inference. LFM2 challenges that assumption. The fashions carry out competitively throughout reasoning, instruction following, multilingual duties, and RAG—whereas concurrently attaining substantial latency beneficial properties over different open small-model households.
For CIOs and CTOs finalizing 2026 roadmaps, the implication is direct: small, open, on-device fashions at the moment are sturdy sufficient to hold significant slices of manufacturing workloads.
LFM2 is not going to exchange frontier cloud fashions for frontier-scale reasoning. But it surely affords one thing enterprises arguably want extra: a reproducible, open, and operationally possible basis for agentic programs that should run anyplace, from telephones to industrial endpoints to air-gapped safe services.
Within the broadening panorama of enterprise AI, LFM2 is much less a analysis milestone and extra an indication of architectural convergence. The longer term will not be cloud or edge—it’s each, working in live performance. And releases like LFM2 present the constructing blocks for organizations ready to construct that hybrid future deliberately slightly than unintentionally.
