Liquid AI's LFM3 unifies text, vision, audio, and decisions

Liquid AI unveiled LFM3, an open weight family of six models built on a common foundation. The goal is clear: to run specialized models on a single GPU, even locally, with video, voice, and calibrated decisions.

Liquid AI presented LFM3 on October 27, 2026, a new generation of open models designed for specialization rather than the usual catch-all approach. The family combines text, vision, audio, omni, and decision on the same base, with models around 2 billion parameters, to run on a single GPU and, according to the company, directly where the application runs. This topic matters for model publishers as well as integrators: LFM3 aims to reduce the gap between capacity, latency, and local deployment.

Une base commune, six variantes spécialisées

Each LFM3 model derives from LFM3-1.7B-Base, a textual foundation presented as the starting point for the entire range. Liquid AI then offers LFM3-1.7B for text, LFM3-VL-2B for text, image, and video, LFM3-Audio-2B for text and audio, LFM3-Omni-2B for visual and sound inputs, then d1-2B and d1-omni-2B for calibrated responses rather than generated text. The idea is simple: no longer force a single model to do everything, which sometimes avoids getting an elegant but off-target answer.

The first building block, LFM3-1.7B, is announced with 1.73 billion parameters, a context window of 256,000 tokens, a vocabulary of 128,000 tokens, and support for 16 languages, according to the model sheet cited in the sources. Liquid AI also states that the base and fine-tuned models are available as open weight on Hugging Face, in the Liquid AI Playground, and via the company's API.

Un nouveau socle technique pour gagner en capacité par paramètre

The core of LFM3 relies on a tri-hybrid architecture: 13 layers of short convolutions, 5 layers of sliding window attention, and 6 layers of global attention. The convolutions increase to a kernel size of 7, compared to 3 on LFM2. Added to this are gated attention per head, partial RoPE, and so-called sandwich normalization. RoPE, or rotary positional embedding, is a token positioning mechanism; here, Liquid AI modulates it to better handle long context and deployments on constrained hardware.

This architecture comes from research designed with a hardware loop in a real environment, extended to 14 combinations of terminals and backends, from smartphones to Raspberry Pi 5, including laptops, cloud CPUs, GPUs, and Qualcomm Hexagon NPU. The company claims that LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters, while being about 25% faster with 8K tokens in context. At 4-bit precision, the text model fits in about 1 GB of memory.

Vision, audio et vidéo : le pari du tout-en-un sans tout alourdir

The vision branch, LFM3-VL-2B, now accepts video input. Liquid AI says it improves on 62 visual benchmarks, with an average rising from 57.6 to 71.4 compared to LFM2.5-VL-3B, while remaining under 2 billion parameters. The company also highlights an internal vision encoder of 164 million parameters, said to consume three times less computation than SigLIP2 in previous LFM VL models, and 2.2 times fewer FLOPs than the Qwen3.5-2B vision encoder according to its measurements.

On the audio side, LFM3-Audio-2B supports 18 languages and claims better speech synthesis than its predecessor, with a UTMOSv2 score of 3.83 versus 3.22 for LFM2.5-Audio-1.5B. Its fully causal audio encoder would allow a time to first token of up to 50 ms, helping to start responding while the person is still speaking. LiquidCodec, the new audio tokenizer, generates speech at about 87 times real speed on a single CPU thread on Apple M4 Max. For a lab model, the cost remains reasonable; for a product, this detail often gets a prototype out of the drawer.

Les modèles de décision changent de logique

With d1-2B and d1-omni-2B, Liquid AI no longer aims to produce a long response but a structured decision. The models receive a situation — support ticket, invoice, application screen, photo, or voice note — then answer several questions in a single pass, with a calibrated probability for each answer. They can also point to objects, read values in documents, estimate a number with a range, or select all applicable options.

According to sources, d1-2B replaces d1-3B with one third fewer parameters and scores 75.5 versus 64.0 on ImajevBench. The interest of this format is clear for business uses: one does not expect well-turned text, but an output exploitable by an application, a ticket, or an automation chain.

Ce que Liquid AI cherche à vendre, au fond

The launch of LFM3 tells a fairly clear strategy: offer smaller models, easier to adapt and less dependent on heavy cloud. Liquid AI says each model can be fine-tuned on a single GPU, with notebooks for text, vision, audio, and omni, from SFT to GRPO including DPO and continued pre-training. The company also claims a launch partner ecosystem including Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm via Qualcomm AI Hub and GenieX, and Vercel, while Mercedes-Benz and Shopify participated in upstream testing according to the publication.

For product teams, the challenge is less the number of benchmarks than the combination of specialization, execution cost, and operational sovereignty. If Liquid AI's figures hold up outside paper, LFM3 may interest embedded uses, from customer support to embedded systems in robotics, including local agents. The usual caution remains: announced performances vary depending on tasks, hardware, and evaluation contexts, so [to verify] on your own data before putting the GPU in the attic.

Key points

  • LFM3 was announced on October 27, 2026, by Liquid AI.
  • Six models share the same backbone LFM3-1.7B-Base.
  • LFM3-1.7B features 1.73 billion parameters and 256K tokens.
  • LFM3-VL-2B adds video input, a first for LFM.
  • LFM3-Audio-2B supports 18 languages.
  • d1-2B replaces d1-3B with one third fewer parameters.

By the numbers

  • 35% — fewer parameters for LFM3-1.7B compared to LFM2.5-2.6B.
  • 22T tokens — pre-training of LFM3-1.7B-Base, according to Liquid AI.
  • 62 benchmarks — visual comparison base for LFM3-VL-2B.
  • 50 ms — time to first token announced for audio.
  • 96.2% — average success claimed in robotics on LIBERO.

To read