Liquid AI's LFM3 ships six 2B-class models for phones, laptops and robots

Update (2026-10-29): Liquid AI has followed the LFM3 family with LFM3’s successor listing on the same Product Hunt page, now described as “Six 2B-class models for phones, laptops and robots” — the identical six-model, three-device framing the article examines. The listing confirms the family strategy is continuing rather than being consolidated into a single flagship, which is the outcome the article flagged as the thing to watch. It does not, however, supply the independent on-device evals, per-variant parameter counts, quantization schemes, or thermal/latency numbers the piece called for, so the core open question — whether the six-model catalog earns its differentiation on real hardware — remains unanswered by this item.

Liquid AI has posted LFM3, a family of six 2B-class language models built for phones, laptops and robots, to its Product Hunt page. The framing is the story: six models, one size class, three device categories. Liquid is not chasing the frontier. It is shipping a product line for the hardware people already own, and it wants the lineup to read like a catalog rather than a single flagship.

That is a real strategic choice, and it deserves a sharper look than the launch copy invites. Six models at roughly 2B parameters is not one model with six checkpoints. It implies six different jobs, each tuned for a device constraint or a task family, and the packaging matters as much as the weights. A robotics deployment and a phone keyboard do not want the same model, even if both fit in 2B parameters. Liquid is selling that differentiation as the product.

Why six models, not one

The 2B-class tier is where on-device inference stops being a demo. A model in that range can run in a few gigabytes of memory, quantized, on a modern phone or a laptop’s integrated GPU, and it can hold a conversation or drive a tool loop without a round trip to a data center. The constraint is not raw capability. It is latency, memory, thermal budget and the cost of every token. Liquid’s bet is that a family of narrow models, each fitted to a device class, beats one general model squeezed onto everything.

That is a defensible position, and it is the opposite of what the big labs are doing. OpenAI, Anthropic and Google DeepMind keep pushing capability up and cost down at the frontier, then distilling smaller versions for the edge. Liquid is starting from the edge and asking what the device can afford. The two approaches can coexist, but they produce different products. A distilled frontier model is a compromise. A model designed for a phone from the first training run is not.

The MIT spinout lineage matters here. Liquid’s earlier work leaned on so-called liquid neural networks and state-space methods, and the company has been public about wanting architectures that scale down cleanly rather than up expensively. LFM3 is the commercial test of that claim. If the six-model family holds up on real devices, it validates the architecture story. If the models feel like generic small LLMs with device labels stapled on, the thesis weakens fast.

The device categories are the tell

Phones, laptops and robots are three very different deployment targets, and grouping them under one 2B-class banner is doing a lot of work. A phone model lives inside an app sandbox with a hard memory ceiling and a battery it cannot drain. A laptop model can borrow more RAM and a fan. A robot model has to run under a control loop where a 200ms stall is a physical problem, not a UX annoyance.

Shipping one family across all three suggests Liquid has built a shared training and evaluation pipeline with per-target tuning, rather than three separate research programs. That is the operationally interesting part. If the same base recipe produces a phone model, a laptop model and a robot model that each clear their device’s bar, Liquid has something closer to a manufacturing process than a research project. That is what a model company needs to be a business rather than a paper.

{/* TODO: verify the specific parameter counts, quantization schemes, and per-device memory footprints for each of the six LFM3 variants — the source listing describes them only as “2B-class.” */}

The economics Liquid is betting on

Small models are a margin play as much as a capability play. A 2B model that runs on a phone costs the vendor almost nothing per inference after the download. A frontier API call costs money every time. For any workload that can tolerate a small model, the on-device version is dramatically cheaper to operate, and it keeps the data on the device, which is increasingly a selling point rather than an afterthought.

That economics is why the small-model tier has gotten crowded. Microsoft has pushed Phi, Google has Gemma, Meta has its smaller Llama variants, and a long tail of startups ships sub-3B models tuned for specific tasks. Liquid’s differentiation cannot just be “small.” It has to be “small and better at the job,” and the six-model framing is how it tries to make that case concrete. A general small model competes on benchmarks. A task-tuned family competes on fit.

The risk is fragmentation. Six models means six evaluation stories, six sets of documentation, six support surfaces, and six chances for one of them to underperform and drag the brand. A single strong model is easier to market and easier to defend. Liquid is trading that simplicity for coverage, and the market will decide whether the coverage is worth the overhead.

Liquid is not chasing the frontier. It is shipping a product line for the hardware people already own.

What this means for builders

For anyone building on-device AI, the LFM3 lineup is worth a look precisely because it is not a single general-purpose model. If Liquid has genuinely tuned separate variants for phone, laptop and robot constraints, that saves a builder the work of squeezing one model into three different boxes. The question to ask is not “how good is LFM3” but “which of the six is tuned for my device, and does it beat the general small model I would otherwise distill.”

The bigger signal is directional. The small-model tier is where the next round of AI product competition will be fought, because that is where the cost per inference collapses and where privacy stops being a compliance checkbox. Liquid’s six-model bet is an early, specific answer to the question every on-device builder is asking: do I tune one model per device, or do I find a family that already did the tuning for me. LFM3 is Liquid’s argument that the family wins. The next few months of real deployments will say whether the argument holds.

The thing to watch is not the launch. It is the first wave of independent evals on actual phones and robots, where the thermal and latency numbers are real and the marketing language falls away. That is where a six-model family either earns its catalog or gets quietly consolidated into one.