# Liquid AI releases six 2B-parameter open-weight models in LFM3 family

Digest AI · Generative AI & Models · published 2026-10-27T17:37:15Z · updated 2026-10-27T19:12:50Z

Canonical: https://digestai.news/story/liquid-ai-releases-six-2b-parameter-open-weight-models-in-lfm3-family

## Summary

Liquid AI launched the LFM3 family on October 27, 2026, introducing six open-weight models with 2 billion parameters each. The lineup includes text, vision, audio, omni, and two decision models, all branching from a shared LFM3-1.7B-Base architecture. The text model supports 16 languages, a 256K-token context window, and a 128,000-token vocabulary, while the vision model accepts video input—a first for Liquid AI—and the omni model handles text, images, video, and audio. Decision models return calibrated answers rather than generated text, with d1-2B replacing the previous d1-3B model with fewer parameters and improved performance on ImajevBench (75.5 vs. 64.0). All models are open-weight and available on Hugging Face, the Liquid AI Playground, and API, under the LFM Open License v1.0. Liquid AI claims LFM3-1.7B outperforms its predecessor LFM2.5-2.6B with 35% fewer parameters, matching LFM2-1.2B on short prompts while running 25% faster with 8K tokens in context. The company also highlights efficiency gains in vision and audio models, with the vision encoder using 3x less compute than previous versions and the audio model supporting 18 languages with real-time speech generation on a single CPU thread. Liquid AI’s tri-hybrid architecture—combining gated convolutions, sliding-window attention, and global attention—was optimized via hardware-in-the-loop testing across 14 device combinations, including phones, laptops, and cloud platforms. The models were pre-trained on 22T tokens, with final stages focusing on agentic traces and tool calls. Liquid AI reported performance gains across benchmarks, including WebArena-Verified (38.42 vs. 3.02) and vision benchmarks (71.4 vs. 57.6), though it notes the text model is not recommended for agentic coding or complex reasoning due to its size. The company also demonstrated fine-tuning applications, such as defect classification in wind-turbine blades and robotics models for orchestration, navigation, and manipulation, with success rates up to 96.2% in specific tasks. Launch partners include Ambarella, AMD, Intel, NVIDIA, and others, while Mercedes-Benz and Shopify participated in pre-release testing.

## Key points

- Liquid AI released six 2B-parameter open-weight models in the LFM3 family, including text, vision, audio, omni, and decision models
- LFM3-1.7B-Base supports 16 languages, 256K-token context, and 128K-token vocabulary, with a tri-hybrid architecture optimized for efficiency
- Vision model LFM3-VL-2B accepts video input and outperforms larger models like Gemma 4 12B on benchmarks with fewer parameters

## Why it matters

The LFM3 family expands open-weight options for developers and researchers, offering multimodal capabilities at lower parameter counts than competitors. Its efficiency gains and hardware support could lower barriers for deployment on edge devices, while decision models with calibrated outputs may appeal to industries needing reliable, structured responses.

## Sources

1. [Liquid AI Launches LFM3 Family of Six Open-Weight 2B-Class Models](https://unite.ai/liquid-ai-launches-lfm3-family-of-six-open-weight-2b-class-models) (Unite.AI, 2026-10-27)
2. [LiquidAI/LFM3-Omni-2B-GGUF](https://huggingface.co/LiquidAI/LFM3-Omni-2B-GGUF) (huggingface.co, 2026-10-27, primary source)
3. [LiquidAI/LFM3-1.7B · Hugging Face](https://huggingface.co/LiquidAI/LFM3-1.7B) (huggingface.co, 2026-10-27, primary source)

## Cite

Digest AI, "Liquid AI releases six 2B-parameter open-weight models in LFM3 family", 27 October 2026, https://digestai.news/story/liquid-ai-releases-six-2b-parameter-open-weight-models-in-lfm3-family

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/liquid-ai-releases-six-2b-parameter-open-weight-models-in-lfm3-family.json