When Liquid AI, a startup fdeveloped by computer scientists at MIT in 2023introduced Series 2 of Liquid Foundation Models (LFM2) in July 2025The premise was elementary: deliver the market’s fastest core models on devices using a modern “seamless” architecture, with training and inference performance that made tiny models a stern alternative to enormous, cloud-only language models like OpenAI’s GPT series and Google’s Gemini.
The first release featured dense checkpointing at 350M, 700M, and 1.2B, a hybrid architecture heavily focused on compact gated convolutions, and benchmark results that placed LFM2 ahead of similarly sized competitors such as Qwen3, Llama 3.2, and Gemma 3 in both quality and CPU throughput. The message for enterprises was clear: real-time privacy-preserving AI on phones, laptops and vehicles no longer required sacrificing capabilities due to latency.
Within a few months of that launch, Liquid expanded LFM2 into a broader product line—adding task- and domain-specialized variants, a tiny video processing and analytics model, and an edge-centric deployment stack called LEAP—and positioned these models as the control layer for on-device and on-premises agent systems.
Now with publication of the detailed 51-page LFM2 technical report on arXivthe company goes a step further: it makes public the architecture discovery process, training data mix, distillation goal, curriculum strategy, and post-training workflows underlying these models.
Unlike previous open models, LFM2 is based on a repeatable recipe: a hardware-in-the-loop search process, a training program that compensates for smaller parameter budgets, and a post-training process tailored to follow instructions and exploit tools.
Instead of simply offering scales and an API, Liquid publishes a detailed blueprint that other organizations can exploit as a reference for training their own tiny, high-performance models from scratch, tailored to their own hardware and implementation constraints.
A family of models designed based on real-world constraints, not GPU labs
The whitepaper begins with a premise that is well known to enterprises: real-world AI systems reach their limits long before benchmarks reach their performance limits. Latency budgets, memory peaks, and thermal throttling define what can actually run in a production environment – especially on laptops, tablets, standard servers, and mobile devices.
To address this issue, Liquid AI performed architecture searches directly on target hardware, including Snapdragon mobile SoCs and Ryzen laptop processors. The result is a consistent result across all sizes: a minimal hybrid-dominated architecture gated compact convolution blocks and a tiny number note regarding group queries (GQA) layers. This design was repeatedly chosen over more exotic hybrids of linear attention and SSM because it provided a Pareto profile with better quality and latency in real device conditions.
This matters to enterprise teams in three ways:
-
Predictability. The architecture is elementary, performance-efficient and stable for models ranging from 350M to 2.6B.
-
Operational portability. Dense and MoE variants share the same structural backbone, simplifying deployment for mixed fleets of equipment.
-
Can be used on the device. Pre-filling and decoding throughput on processors outperforms comparable open models by approximately two times in many cases, reducing the need to offload routine tasks to cloud inference endpoints.
Rather than optimizing for academic news, the report reads as a systematic attempt to design models that companies can achieve actually send.
This is noteworthy and more practical for enterprises operating in an industry where many open models silently assume access to clusters containing multiple H100s during inference.
Training scheme adapted to behaviors crucial for the company
LFM2 adopts a training approach that compensates for the smaller scale of its models with structure rather than brute force. Key elements include:
-
Initial 10-12T token training and additional The middle phase of training in the context of 32,000which extends the useful context window of the model without increasing computational cost.
-
AND separated purpose of distilling Top-K knowledge this avoids the instability of the standard KL distillation when teachers only report partial logits.
-
AND three-stage post-training sequence– SFT, Normalized Length Preference Matching and Model Linking – designed to provide more hearty instruction execution and tool exploit behavior.
What’s crucial for enterprise AI developers is that LFM2 models behave less like “little LLMs” and more like hands-on agents that can follow structured formats, follow JSON schemas, and manage multi-step chat flows. Many open models of similar size fail not because of a lack of reasoning capacity, but because of a frail adherence to instructional templates. LFM2’s post-workout prescription directly targets these pointed edges.
In other words: Liquid AI optimized for tiny models operational reliabilitynot just the scoreboards.
Multimodality designed for device limitations, not laboratory demonstrations
The LFM2-VL and LFM2-Audio variants reflect another change: multimodality built around token effectiveness.
Instead of embedding a huge vision transformer directly into the LLM, LFM2-VL connects the SigLIP2 encoder via a connector that aggressively reduces the number of visual tokens via PixelUnshuffle. High-resolution inputs automatically trigger vigorous tiling, so token budgets can be controlled even on mobile devices. LFM2-Audio uses a forked audio track – one for embedding, one for generating – supporting real-time transcription or speech-to-speech on modest processors.
For enterprise platform architects, this project points to a practical future in which:
-
document understanding occurs directly at endpoints such as field devices;
-
audio transcription and speech agents run locally to ensure privacy policy compliance;
-
multimodal agents operate within fixed-latency envelopes without streaming data outside the device.
The through line is the same: multimodal capabilities without the need for a GPU farm.
Search models built for agent systems, not legacy search
LFM2-ColBERT extends overdue interaction search functionality to an area tiny enough for enterprise deployments that require multilingual RAG, without the need for specialized vector database accelerators.
This is especially crucial as organizations begin to organize fleets of agents. Rapid local retrieval – running on the same hardware as the inference model – reduces latency and provides management benefits: documents never leave the boundaries of the device.
In summary, the VL, Audio and ColBERT variants show the LFM2 as a modular system rather than a single model.
The emerging design of hybrid enterprise AI architectures
In all variants, the LFM2 report implicitly sketches what the future enterprise AI stack will look like: hybrid on-premises cloud orchestrationwhere tiny, swift models running on devices handle time-critical tasks of perception, formatting, invoking tools and evaluation, while larger models in the cloud offer advanced reasoning when needed.
Several trends converge here:
-
Cost control. Running routine inferences locally helps avoid unpredictable billing in the cloud.
-
Latency determinism. TTFT and decoding stability matter in agent workflows; on the device eliminates network vibrations.
-
Governance and compliance. Local execution simplifies the handling of PII, data storage, and auditability.
-
Resistance. Agent systems gradually degrade if the cloud path becomes unavailable.
Enterprises implementing these architectures will likely treat tiny models on devices as the “control plane” for agentic workflows, while enormous models in the cloud will act as on-demand accelerators.
LFM2 is one of the clearest open source foundations for this control layer to date.
Strategic takeaway: On-device AI is now a design choice, not a compromise
For years, organizations building AI capabilities have assumed that “true AI” requires inference in the cloud. LFM2 challenges this assumption. The models perform competitively on reasoning, instruction execution, multilingual tasks, and RAG, while achieving significant latency gains compared to other open families of tiny models.
For CIOs and CTOs finalizing their 2026 roadmaps, the implications are immediate: tiny, open models on the device are now sturdy enough to withstand a significant portion of production loads.
LFM2 will not replace boundary cloud models for frontier scale reasoning. But it offers something that enterprises probably need more of: a repeatable, open, and operationally viable foundation agent systems that need to run anywherefrom phones to industrial endpoints to secure air-insulated facilities.
In the expanding enterprise AI landscape, LFM2 is less of a research milestone and more of a sign of architectural convergence. The future is not a cloud or a shore – it is both, working together. Releases like LFM2 provide building blocks for organizations prepared to build this hybrid future with purpose, not accident.