As of: 2026-11-13 13:49 CET.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/m/liquid-lfm3-1-7b/

# LFM3-1.7B

Liquid AI, LFM3 family. Status: current. Type: language model. Released: 2026-10-27. Training cutoff: not published.

## How to call it

| Provider | Model id | Endpoint | Docs |
|---|---|---|---|
| Hugging Face (text model; also LFM3-1.7B-Base, GGUF and MLX builds) | `LiquidAI/LFM3-1.7B` |  | https://huggingface.co/LiquidAI/LFM3-1.7B |
| Hugging Face (siblings on the same backbone) |  |  | https://huggingface.co/LiquidAI |

## Pricing

- Price not published.
- Last checked: 2026-10-29

## Specs

- Input: text
- Output: text
- Context window: 256,000 tokens
- Max output: not published
- Open weights: yes
- Licence: lfm1.0

## What it has not seen

- No published training cutoff. Counted from its release (2026-10-27): at least 365 AI events since (69 major, 6 historic), as of 2026-11-13.
- Everything since September 2026: https://postcutoff.com/since/2026-09/

## What stands out

- Small on-device agent model: 1.7B tri-hybrid model (13 gated short convolutions, 5 sliding-window and 6 global attention layers), ~1 GB at 4-bit, 256K context; Liquid reports WebArena-Verified 38.42 vs 3.02 for LFM2.5-2.6B and says it beats Qwen3.5-9B on long browser and business-app tasks (vendor numbers). Source: https://huggingface.co/blog/LiquidAI/lfm3
- One backbone for every modality: Six 2B-class open models branch from LFM3-1.7B-Base: text, LFM3-VL-2B (first LFM with video input), LFM3-Audio-2B (18 languages, speech in and out), LFM3-Omni-2B (Liquid's first omni model) and the d1-2B / d1-Omni-2B decision models. Source: https://huggingface.co/blog/LiquidAI/lfm3

## Notes

Released Oct 27, 2026 under the LFM Open License v1.0 (Hugging Face 'lfm1.0'). Pre-trained on 22T tokens with 3.5T tokens of mid-training on agentic traces; post-trained for agent harnesses such as Hermes Agent, Pi and OpenClaw; fine-tunes on one GPU. Benchmarks are Liquid's own.

Run with a recent `transformers` (the vision, audio and omni siblings need 5.20–5.21 and `trust_remote_code=True`); see the model
cards for llama.cpp and MLX builds.

Model page: https://postcutoff.com/m/liquid-lfm3-1-7b/