LFM3 review: Liquid AI's six open-weight 2B multimodal models for on-device text, vision, audio and agents — honest scores, strengths, limits and fit.
LFM3 is a strong, genuinely useful release — six open-weight 2B-class multimodal models that run privately on a phone or laptop and punch above their size on agentic and tool-use tasks. It is a developer and infrastructure product, not a creator app: there is no interface for making finished content, and the self-reported benchmarks need independent confirmation. Judge it as a building block you fine-tune and embed, and it earns its score; expect it to write your blog and post your Reels and you are using the wrong layer of the stack.
Most of the attention in AI goes to models that get bigger. LFM3 — announced by Liquid AI on October 27, 2026 at liquid.ai/blog/lfm3 — goes the other way. It is the third generation of Liquid Foundation Models, and the first release is a family of six open-weight models, all in the 2B-parameter class, built from one shared backbone and tuned to run on devices rather than in a datacenter. The pitch, in Liquid's words, is "every modality for specialized intelligence": a small base you fine-tune into whatever narrow job you need.
This review judges LFM3 on what it actually is. That matters because the name sits next to models you might assume it competes with, and it does not compete with most of them. It is not a chatbot you open and talk to, and it is not a content tool. It is a set of weights — text, vision, audio, omni, and two agent-focused "decision" models — that a developer downloads, fine-tunes, and embeds. Reviewed on those terms, the release is impressive; reviewed as something a non-technical creator can pick up and publish with, it is the wrong tool entirely.
I run Kompozy, which is a content engine, not a foundation model, so there is no rivalry here to color the scoring — LFM3 operates a layer below anything Kompozy does. That lets me be blunt in both directions: Liquid AI has shipped something genuinely clever, and it will do nothing on its own for the creator who just wants a week of posts. Both of those are true.
Every figure below is grounded in Liquid AI's launch material and the launch-day coverage. Where the numbers are self-reported — and most of the benchmark claims are — this review says so rather than presenting them as settled.
LFM3 is a family of six open-weight models in the 2B-parameter class, all branched from a single backbone pre-trained on 22 trillion tokens. The lineup spans five modalities: a text model (LFM3-1.7B) with a 256K-token context window, a vision model that reads images and video, an audio speech model with speech in and out, an "omni" model that ingests text, vision, and audio together and returns both text and speech, and two "decision" models specialized for agentic action. Architecturally it is a hybrid — across 24 layers it mixes gated short convolutions, sliding-window attention, and a few global-attention layers — which is how it holds a small memory footprint while reaching a long context. The defining trait is efficiency on-device. Liquid AI says the text model fits in roughly 1 GB at 4-bit precision, can be fine-tuned on a single GPU, and decodes 1.8 to 2.8 times faster than comparable small models it tested locally, with a much smaller key-value cache. The weights are available on Hugging Face, the Liquid Playground, and the Liquid AI API, under an "LFM Open License" that launch coverage describes as free for commercial use below roughly a $10M annual-revenue threshold — a term to confirm on the model card before building commercially.
LFM3 is for developers, ML engineers, and technically staffed teams who want a private, low-latency, multimodal model they can run on a phone, PC, or edge device and fine-tune on their own data. It is a strong fit for on-device transcription, image understanding, narrow agentic tasks, and privacy-sensitive processing where sending data to a cloud API is a non-starter. It is not for a creator, marketer, or small business that wants to open an app and produce finished, published content — there is no such interface here, and assembling one around the raw weights is a software project, not an afternoon.
| Dimension | Score | Why |
|---|---|---|
| On-device efficiency | 4.7 / 5 | Fitting a capable multimodal model in ~1 GB at 4-bit with a far smaller KV cache and 1.8–2.8× faster local decode is the headline strength, and it is credible for a 2B-class model. |
| Multimodal breadth | 4.5 / 5 | Text, vision, audio, and a true omni model that takes all three and returns text and speech — unusually complete coverage from one small backbone. |
| Agentic / tool-use performance | 4.3 / 5 | Liquid AI reports it beating much larger models on function-calling and web-automation suites; strong if it holds up, but the figures are self-reported. |
| General knowledge & reasoning | 3.3 / 5 | Liquid AI concedes larger models still lead on broad knowledge, long-context recall, and agentic coding — expected for 2B parameters, but a real ceiling. |
| Openness & licensing | 3.9 / 5 | Open weights on Hugging Face are a genuine plus; the revenue-capped commercial license is more restrictive than a pure Apache/MIT release. |
| Fine-tuning & customization | 4.4 / 5 | Single-GPU fine-tuning on an open base makes a custom, private model realistic for a small technical team. |
| Ease of use for non-developers | 2.2 / 5 | There is no product surface for a non-technical user — it is weights and endpoints; a Playground exists, but turning it into a workflow requires engineering. |
| Documentation & ecosystem | 3.8 / 5 | Docs for local running and fine-tuning plus llama.cpp support and launch partners (AMD, Ambarella, Qualcomm) are a solid start for a brand-new family. |
There is no subscription to evaluate in the usual sense: LFM3's weights are open and free to download from Hugging Face, and you can also reach the models through the Liquid Playground and API. The real cost is not a sticker price but engineering and compute — the time to integrate, fine-tune, host, and maintain the model, plus whatever hardware you run it on. For a team that already builds software, "free weights you run on a single GPU" is a strong deal. For anyone without that capacity, the effective cost is high because the model does nothing turnkey.
The licensing nuance matters. Launch coverage describes an "LFM Open License" that is free for commercial use below roughly a $10M annual-revenue threshold, which is more permissive than a research-only license but more restrictive than Apache 2.0 or MIT. Larger companies may owe a commercial arrangement; smaller teams likely fall under the free tier. Confirm the exact terms on the official model card before shipping anything commercial.
Compared with paying per token for a hosted frontier model, running LFM3 on-device can be dramatically cheaper at volume for the narrow tasks it is good at — and free of the latency and privacy cost of a cloud round-trip. That is the whole economic argument for a small on-device model, and LFM3 makes it well, provided you have the skills to realize it.
| Use case | Fit | Why |
|---|---|---|
| On-device transcription and audio processing | Strong | The audio and omni models handle speech locally, which is exactly where a private, low-latency small model shines. |
| Embedding a private multimodal model in your own app | Strong | Open weights, single-GPU fine-tuning, and edge/on-device design make this the intended use. |
| Narrow agentic tasks and tool calling | Strong | Liquid AI positions LFM3 around function-calling and automation, with benchmarks (self-reported) to match. |
| Building a custom brand model on your own data | OK | Single-GPU fine-tuning makes it feasible, but you still need the engineering to train, host, and evaluate it. |
| Broad knowledge Q&A or deep reasoning | Weak | At 2B parameters it trails larger models on general knowledge and reasoning by Liquid AI's own admission. |
| A creator producing finished social posts | Weak | There is no interface for making or publishing content — this is a model, not a content tool. |
| Captioning, reframing, and multi-platform publishing | Weak | None of this exists in LFM3; it is a separate workflow handled by an engine like Kompozy. |
This is less a comparison than a clarification of layers. LFM3 is a foundation-model family — infrastructure a developer builds on. Kompozy is a finished content engine a creator uses. They do not overlap: if you dropped LFM3 into Kompozy's stack, it would be one more model behind the scenes, not a substitute for the product. So "LFM3 vs Kompozy" only makes sense as a question about which layer solves your problem.
If your problem is "I want a private, low-latency multimodal model I can fine-tune and embed," LFM3 is a real answer and Kompozy is irrelevant to it. If your problem is "I want a week of on-brand video, carousels, blogs, and newsletters published across every platform," LFM3 does none of that and Kompozy does all of it — generating across 18 formats and publishing across eight social platforms plus blog and email, held to one voice by a Persona Brief, scheduled and shipped on Autopilot. The honest framing: LFM3 is a brilliant component, Kompozy is the application, and most people searching for one actually want the other.
For developers and technical teams who want a private, efficient, multimodal model to fine-tune and run on-device, yes — the efficiency and multimodal breadth are genuinely strong for a 2B-class family. For a non-technical creator or marketer who wants finished content, no: LFM3 has no interface for making or publishing posts, and building one around the raw weights is a software project.
On-device, privacy-sensitive, low-latency tasks and narrow agentic work — transcription, image and video understanding, speech, and tool/function calling — where its small footprint and local speed matter more than frontier-scale knowledge. Liquid AI reports it beating much larger models on agentic benchmarks, though those figures are self-reported.
Broad general-knowledge Q&A, deep reasoning, and agentic coding, where larger models still lead by Liquid AI's own admission. It also does nothing end-to-end for content — no captioning, reframing, formatting, scheduling, or publishing.
The weights are open and free to download from Hugging Face, with access also via the Liquid Playground and API. Launch coverage describes an "LFM Open License" that is free for commercial use below roughly a $10M annual-revenue threshold; confirm the exact terms on the model card before shipping commercially.
They are the nearest open, small-model alternatives, and Liquid AI benchmarks LFM3 against them directly — claiming strong agentic and efficiency results while conceding larger models lead on knowledge. Treat the launch-day comparisons as self-reported until independent evaluations land.
Not on its own. LFM3 generates and understands content but has no path to finished, published posts. Turning model output into captioned, reframed, scheduled content across platforms is a separate job handled by a content engine like Kompozy.
Treat them as promising but unconfirmed. The launch figures are self-reported by Liquid AI, and at least one outlet flagged specific results as not yet independently verified. They point to a capable small model; wait for third-party evaluations before treating any single number as settled.