📊 Browse AI model rankings on AI Rank

📱 Get the AI Rank app

LFM3 (Liquid AI LFM3) is Liquid AI’s October 27, 2026 generation of Liquid Foundation Models. The company blog Introducing LFM3 describes a multimodal family—text, vision, audio, omni, and decision—on one shared backbone in the ~2B parameter class, with open weights on Hugging Face, the Liquid Playground, and the Liquid AI API.

LFM3-1.7B and what shipped

Highlights from the official post (checked October 28, 2026):

  • New tri-hybrid architecture (gated short convolutions + sliding-window attention + global attention) after 22T pre-training tokens.
  • First release: six 2B-class models branched from LFM3-1.7B-Base.
  • LFM3-1.7B text model: 256K context; fits about 1 GB at 4-bit per the post.
  • Open weights; docs cover local run and fine-tune entry points.
  • Claim vs prior gen: LFM3-1.7B outperforms LFM2.5-2.6B while using 35% fewer parameters (Liquid’s stated comparison).

Hugging Face collection: LiquidAI/lfm3 (updated 2026-10-27). Example repos called out there include LiquidAI/LFM3-1.7B-Base and LiquidAI/LFM3-1.7B (with GGUF / ONNX / MLX / FP8 / NVFP4 siblings linked in collection notes).

Family map (official names)

From the blog’s modality sections:

  • Text — LFM3-1.7B: local agents, tool use, long context (256K). Post-trained for harnesses named in the article (Hermes Agent, Pi, OpenClaw, among others).
  • Vision — LFM3-VL-2B: images + video, step-by-step thinking, spatial/embodied reasoning; new in-house vision encoder (Liquid says 3× less compute than the prior SigLIP2 path).
  • Audio — LFM3-Audio-2B: text/audio in; text and/or speech out; 18 languages for speech tasks; tool calling from speech.
  • Omni — LFM3-Omni-2B (~2.2B): vision + audio + language together (spoken VQA, audio-visual video, multimodal tool calls).
  • Decision — d1-2B (from VL) and d1-Omni-2B (from Omni): answer many typed questions in one forward pass with calibrated probabilities; Point / Read / Estimate / Pick answer types described in the post.

Architecture notes (blog-level)

Liquid publishes a 24-layer mix: 13 convolutions (kernel 7), 5 SWA (256-token window), 6 global attention; head-wise gated attention; partial RoPE on global layers; sandwich norm; 128K vocabulary shared with LFM2.5-8B-A1B. Hardware-in-the-loop search expanded to 14 device/backend combinations for LFM3. Mid-training’s last 3.5T tokens emphasize curated/agentic data and extend context to 256K.

Local and open-weight entry points

Official “Get Started” points to:

  1. Hugging Face weights (collection above) — download, fine-tune, deploy.
  2. Liquid AI API and Playground.
  3. Native stack callouts: llama.cpp, MLX, vLLM, plus Apple / AMD / NVIDIA / SGLang / Qualcomm paths, and Liquid’s Real-time Inference Pipeline for low-latency serving.

The blog says every model can be fine-tuned on a single GPU and links fine-tuning notebooks for specialization demos (for example wind-turbine inspection on LFM3-VL-2B). We are not pasting unverified shell transcripts here—follow Liquid’s live docs for exact commands.

Day-0 partners named on the post include Ambarella, AMD, Arm, Intel (OpenVINO), NVIDIA, OpenRouter, Qualcomm, and Vercel, among builder tools (Ollama, LM Studio, Unsloth, etc.).

LFM3 vs LFM2 (what Liquid claims)

  • More capability per parameter via the new hybrid recipe; LFM3-1.7B vs LFM2.5-2.6B example above.
  • Multimodal family on one backbone (LFM2 era was less complete on this page’s framing).
  • Long-context speed: with 8K tokens in context, LFM3-1.7B about 25% faster per parameter vs LFM2-1.2B on edge devices (Liquid’s wording).
  • New vision/audio encoders and first omni/decision pairings in this release narrative.

All benchmark tables in the blog (LiveClaw, BFCL, SWE-bench, MMMU-style vision suites, ASR WER, Decision Bench, and more) are Liquid-reported. Cite the post; do not treat them as AI Rank leaderboard rows unless they appear in our own snapshots.

→ Rankings: explore related models on AI Rank models

📱 Follow launches in the AI Rank app

FAQ

What is LFM3?

Liquid AI’s LFM3 foundation-model family (announced October 27, 2026): open-weight ~2B multimodal models for text, vision, audio, omni, and decision workloads.

Are LFM3 weights open?

Yes for the base and post-trained models named in the blog—available on Hugging Face, with API/Playground access as well.

How big are the models?

First release is in the 2B class; the shared backbone is LFM3-1.7B-Base. Omni is listed around 2.2B.

How do I run LFM3 locally?

Start from the LFM3 blog Get Started links and the HF collection; Liquid documents llama.cpp / MLX / vLLM-style paths. Use their current docs for commands—this article does not invent install logs.

Sources

Keep exploring: see live rankings on AI Rank models.

Or 📱 get the AI Rank app.

Prepared with automated editorial assistance from Liquid’s LFM3 post and HF collection metadata on October 28, 2026 (America/Los_Angeles). Benchmarks are vendor-reported; we did not re-run them.