Liquid AI LFM3: Six Open-Weight 2B Models for Text, Vision, Audio and Decisions

Liquid AI has released LFM3, an open-weight family of six models in the roughly 2B-parameter class that cover text, vision with video, audio, omni input and calibrated decisions on a single shared backbone. The company's launch post on Hugging Face calls it a starting point for specialized intelligence you can fine-tune on one GPU. The benchmark numbers are Liquid's own; the models, though, are downloadable now, so you can test the claims yourself.

TL;DR: the LFM3 lineup

LFM3-1.7B-Base is also released for custom post-training. Collection: LiquidAI/lfm3 on Hugging Face.

Why one backbone matters

Most small-model families ship a text model, then bolt on separate vision or audio systems built by different recipes. Liquid's pitch is that all six LFM3 models share one language backbone, LFM3-1.7B-Base, so a team can learn one fine-tuning workflow and apply it across modalities. The vision encoder is a 164M-parameter in-house model, and the audio encoder is fully causal, which suits streaming speech.

This follows the company's earlier on-device work. We covered the previous generation in LFM2.5-2.6B for on-device agents and its Pipette on-device benchmark suite. LFM3 is the bigger step: more modalities, a new layer mix and longer context.

The architecture: a tri-hybrid

According to Liquid, LFM3-1.7B-Base has 24 layers made of:

• 13 gated short convolution layers (kernel size raised from 3 to 7)
• 5 sliding-window attention layers with a 256-token window
• 6 global attention layers

Three other changes from LFM2: head-wise gated attention (a sigmoid gate scales each head's output), partial RoPE (global attention layers rotate only half of each head), and sandwich norm (each block normalizes its input and output). Liquid says the layout was found by hardware-in-the-loop search across 14 device and backend combinations, so it is tuned to real phones and laptops, not only GPUs.

Pre-training used 22T tokens with a 128K vocabulary, followed by 3.5T tokens of mid-training that extend context to 256K. Few attention layers means a smaller key-value cache, which is what makes long context feasible on small devices.

What Liquid claims (vendor-reported)

Efficiency claims: the text model fits in about 1 GB at 4-bit precision, audio time to first token can be as low as 50 ms, and the LiquidCodec decodes at about 87 times real time on one CPU thread on an Apple M4 Max.

Treat these as marketing until reproduced. A jump from 3.02 to 38.42 on a web-agent benchmark is large enough to deserve your own test, and "on par with Gemma 4 12B" depends on which 62 benchmarks you average. These comparisons are against Liquid's own earlier models and a few named rivals, not a neutral leaderboard.

Which LFM3 model for which job

The choice that matters most is whether you need one model or several. With a shared backbone, you can prototype with the text model and then swap in the vision or audio variant without relearning a toolchain.

What people will ask

Is 2B enough for agents? For narrow, well-specified tool calls, often yes, especially after fine-tuning. For open-ended planning, small models still fail more than frontier ones. Liquid's WebArena-Verified result of 38.42 is high for this size by its own account, but web agents are brittle, so test on your own flows.

Why does a hybrid with so few attention layers help? Convolution layers keep a fixed small state, and sliding-window attention only looks at the last 256 tokens. Only six layers attend globally, so the cache that grows with context is small. That is what lets a 256K window sit on a laptop or phone instead of a server.

Can it run on a phone? Liquid's 1 GB figure at 4-bit precision suggests the text model can, and the 50 ms audio time-to-first-token claim is aimed at voice on-device. We have not measured either. Real-world speed depends on the runtime and chip.

How is it different from Gemma or Qwen small models? Mostly architecture and packaging. Gemma and Qwen families are mostly attention-based transformers; Liquid leans on convolution hybrids and ships text, video, audio, omni and decision models together. Whether that beats a tuned alternative on your task is an empirical question, which is why the single-GPU notebooks matter.

What about safety and misuse? Open weights can be fine-tuned by anyone, including to remove guardrails, and the announcement describes no safety evaluation beyond capability results. If you deploy an LFM3 model in a product, add your own input and output filtering and log failures.

A sensible evaluation plan

1. Write 50 to 200 real examples from your product, with expected outputs.
2. Run LFM3 and your current model on them with identical prompts.
3. Record accuracy, latency and peak memory on your target hardware.
4. Fine-tune with Liquid's notebook on a few hundred of your own examples and re-measure.
5. Decide on cost and privacy, not benchmark averages alone.

If the fine-tuned 2B model holds up, you gain offline use, no per-token bill and data that never leaves the device. If it does not, you have a cheap baseline that tells you how much you are paying a frontier model for.

How to try LFM3

1. Pick a model from the collection and read its model card for the license.
2. Install a recent transformers: 5.21.0 or later for LFM3-VL-2B, 5.20.2 or later for Audio and Omni, 5.19 or later for d1-2B. The text model loads through a pipeline call with trust_remote_code set.
3. Use the snippets on each card. To fine-tune, Liquid links Colab notebooks in its cookbook for SFT, DPO, GRPO and continued pre-training across text, vision, audio and omni on a single GPU.
4. For local agents, Liquid points to a guide for harnesses such as OpenClaw, Hermes Agent and Pi in its documentation.

Where small models win

LFM3 fits a trend we have tracked all month: narrow, fine-tuned small models beating bigger ones on a measured task. Coinbase reported that a fine-tuned Qwen3.5 9B beat Opus 4.5 on fraud detection, and an edge alarm triage model came in 49 times cheaper than Claude Opus. If your task is narrow and latency or privacy matters, a 2B model you control can beat a frontier API on cost and speed, even when it loses on general ability.

For Apple hardware, see how Gemma 4 26B-A4B sped up on Macs with MLX and the Edge0 35B on-device model for the competing options.

Limits and open questions

• License unclear in the post. Open-weight is not the same as permissive; read each card.
• Vendor-only numbers. No independent leaderboard result was cited.
• Small-model ceiling. A 2B model will not match large models on broad knowledge or hard reasoning.
• Decision models are niche. The d1 models return calibrated answers, a different output type from chat. Evaluate whether calibration holds on your data.
• Transformers version churn. Required versions are very recent, so pin your environment.

What this means for you

If you ship on-device features, LFM3 gives you one family and one fine-tuning path across modalities, with a text model near 1 GB at 4-bit. Run it against your own test set next to your current model, and compare latency and memory on the target device, not just accuracy. If you only call cloud APIs, the practical takeaway is a cost benchmark: ask whether a small specialised model could handle your highest-volume task.

Related reading

• Liquid AI LFM2.5-2.6B on-device agents
• Liquid AI Pipette on-device benchmark
• Coinbase Qwen3.5 9B fraud agent
• Exowatt fine-tuned Qwen3 vs Claude Opus
• Gemma 4 26B-A4B on Mac with MLX
• Edge0 35B on Apple Silicon

Specs and benchmark claims are from Liquid AI's announcement and accurate as of October 27, 2026.