Model releases · 28 October 2026
Liquid AI's third generation starts small on purpose: six models of around 2B parameters, all branched from one 1.7B backbone, with a text model the lab says fits in about 1 GB of memory at 4-bit.
| Released by | Liquid AI |
|---|---|
| Announced | 27 October 2026 |
| Text model | LFM3-1.7B — 1.73B parameters |
| Vision model | LFM3-VL-2B — 1.73B language model + 164M vision encoder + 25M projector |
| Architecture | Hybrid, 24 layers: 13 gated short convolution, 5 sliding-window attention, 6 global attention |
| Context length | 256K tokens (text model); 128K tokens (vision model) |
| Memory at 4-bit | About 1 GB for the text model, per Liquid AI |
| Pre-training | 22T tokens, including a context-extending mid-training stage |
| Languages | 16: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish |
| Reasoning | Thinking on by default; can be switched off in the chat template |
| Licence | LFM Open License v1.0 — not Apache or MIT |
| Official builds | BF16, FP8, NVFP4, GGUF, MLX 4-bit, ONNX |
| Hosted at | LiquidAI on Hugging Face |
LFM3 is the third generation of Liquid AI's foundation models. The first release is deliberately small: six models of roughly 2B parameters, all branched from a single LFM3-1.7B-Base backbone. There is a general text model, separate vision, audio and omni models, and two “decision” models specialised from the vision and omni ones.
The text model, LFM3-1.7B, is post-trained for agentic use. Liquid AI says it outperforms its own LFM2.5-2.6B with 35% fewer parameters, and that it fits in about 1 GB of memory at 4-bit precision.
The 24-layer stack mixes 13 gated short-convolution layers with 5 sliding-window and 6 global attention layers. Liquid AI says the layout came from a hardware-in-the-loop search, measured on real devices rather than chosen on paper.
The practical effect is on long context. Liquid AI reports that a 128K-token context needs 1.6 GB of KV cache, against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B, measured by Liquid AI in llama.cpp at Q4_0.
LFM3-VL-2B pairs the same 1.73B language model with a 164M-parameter vision encoder. Images are processed at native resolution and video is sampled at 2 frames per second. Liquid AI positions it for reading documents, charts and app screens, finding objects and spatial reasoning, and says it is not recommended for coding, long multi-turn conversations or complex reasoning.
The model cards list 16 languages. Anything outside that list, Turkish for example, is not supported, and answers in those languages should not be expected to hold up.
Like LFM2.5, LFM3 ships under Liquid AI's own LFM Open License v1.0. Read it before redistributing the weights inside a commercial product.
| Runtime | Supported | Notes |
|---|---|---|
| llama.cpp | Yes | Official GGUF from Liquid AI; the lab ran its own on-device measurements in llama.cpp. |
| Ollama | Yes | Ollama announced LFM3 availability on 27 October 2026. |
| MLX | Yes | Official MLX 4-bit build; the vision model runs through MLX-VLM. |
| ONNX Runtime | Yes | Official ONNX export. |
| Vision input | Partial | Only in LFM3-VL-2B, not in the text model. |
Liquid AI's third-generation model family, released on 27 October 2026. The first release is six models of around 2B parameters, led by the 1.7B text model LFM3-1.7B and the vision model LFM3-VL-2B.
About 1 GB at 4-bit precision, according to Liquid AI.
256K tokens for LFM3-1.7B and 128K tokens for LFM3-VL-2B.
The vision model, LFM3-VL-2B, can. The text model LFM3-1.7B cannot.
Sixteen: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian and Polish.
Liquid AI's LFM Open License version 1.0, not Apache 2.0 or MIT.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 28 October 2026