LFM3

LFM3 is Liquid AI's new generation of Liquid Foundation Models (LFMs), announced on 2026-10-27 and built for specialized intelligence on the edge. The first release covers six ~2B-parameter models, all branched from a shared LFM3-1.7B-Base backbone: text, vision, audio and omni models, plus two decision models specialized from the vision and omni variants.

Announcement: liquid.ai/blog/lfm3. Hugging Face blog: Build specialized multimodal models with LFM3. Base and post-trained checkpoints are open-weight on Hugging Face, the Liquid Playground and the Liquid AI API.

Key points

1. One backbone, every modality: pick the 2B-class model for your modality and fine-tune it, instead of running one large general model.
2. New architecture: a tri-hybrid architecture and training recipe; Liquid says LFM3-1.7B beats its previous LFM2.5-2.6B with ~35% fewer parameters. Pretrained on ~22T tokens.
3. Built for the edge: the text model fits in about 1 GB at 4-bit and runs on phones and laptops; 256K context for text.
4. LFM3-VL-2B: averages 71.4 on Liquid's 62-benchmark vision suite (vs 57.6 for LFM2.5-VL-3B) and is reported to match Gemma 4 12B (~6× larger) — vendor-reported numbers.
5. New in-house vision and audio encoders; every model can be fine-tuned on a single GPU.
6. Day-one ecosystem: llama.cpp, MLX, vLLM and SGLang across Apple, AMD, NVIDIA and Qualcomm hardware; also available via partners such as OpenRouter.

Why this radar term

Officially launched, newly named model family inside this window: new liquid.ai sitemap URL /blog/lfm3, an X trending news story and posts from prominent AI accounts (@omarsar0, @xenovacom). The radar previously only tracked the LFM2.5 line.

Get started

1. Read liquid.ai/blog/lfm3 to see what each of the six models is for
2. Download weights from the LiquidAI organization on Hugging Face (e.g. LFM3-1.7B, LFM3-VL-2B) or try them in the Liquid Playground
3. Run locally with llama.cpp (GGUF), MLX on Apple Silicon, or serve with vLLM / SGLang
4. Specialize: follow Liquid's fine-tuning guide to SFT a model on your own data on a single GPU
5. Agents: Liquid documents running LFM3 inside local agent harnesses (e.g. Hermes Agent, Pi); in-browser demos run via Transformers.js + WebGPU

Good fit

• On-device, privacy-first or low-latency applications
• Teams fine-tuning small models into task experts (documents, speech, robot spatial reasoning)

Caveats

• Benchmarks are self-reported; independent evaluations are pending
• Released under the LFM Open License v1.0 — check commercial terms