Liquid AI releases LFM3: six on-device models spanning text, vision, audio and omni

LiquidAI released its next-generation Liquid Foundation Models — the LFM3 series — on October 28, launching six on-device model instances spanning text, vision, audio, and omni-modal capabilities, including both generative and decision-making models. Officially positioned as an on-device suite for "specialized intelligence." Model weights are now open-sourced on Hugging Face, including both base and post-trained versions.

Confirmed

• LFM3 adopts a new tri-hybrid architecture and a new multimodal encoder; per @jacek2023 and @maximelabonne, the 1.7B text model consists of 13 layers of gated short convolutions + 5 layers of sliding-window attention + 6 layers of global attention, 24 layers in total.
• Family members include LFM3-1.7B (text-only, 256K context, 1GB with 4-bit quantization), LFM3-VL-2B (image-text understanding), LFM3-Audio-2B (audio), and LFM3-Omni-2B (omni-modal), ranging from 1.7B to 2B parameters.
• Weights are open-sourced on Hugging Face; @omarsar0 noted both base and post-trained versions are included (the post mentions "seven models," slightly at odds with the official count of six instances).

Unconfirmed

• @FinancialAd1961 relayed official claims that the 1.7B text model ranks first on the on-device intelligence Pareto frontier and that VL-2B rivals GPT-5.6 Luna (also mentioned in a repost by @JosephJacks). Both are vendor self-assessments and lack independent benchmark verification.

Why it matters

LFM3 packs 256K long-context and omni-modal capabilities into under 2B parameters with open weights, targeting local agents and on-device deployment — a direct challenge from the small-model on-device track to Big Tech's closed-source models.

2026-10-28 ~ 2026-10-28 · 14 related posts

• [source] LiquidAI Drops Six LFM3 Edge Models: 1.7B Text With 256K Context Fits in 1GB at 4-bit — jacek2023 · 2026-10-28
• [source] Liquid AI launches LFM3: tri-hybrid architecture, six edge models, all modalities — maximelabonne · 2026-10-28
• [source] LiquidAI releases LFM3 weights on Hugging Face: seven models from 1.7B text to 2B omni — omarsar0 · 2026-10-28
• Liquid AI ships LFM3 on-device small model family; 1.9B vision model rivals 12B-class — mhdfaran · 2026-10-28
• LFM3-1.7B beats its 2.6B predecessor with 35% fewer params, 256K context — JosephJacks_ · 2026-10-28
• Liquid AI launches LFM3: open-weight 2B multimodal family on one backbone — JosephJacks_ · 2026-10-28
• LiquidAI's LFM3 lands on Ollama: ~1GB text model at 4-bit with 256K context — ollama · 2026-10-28
• Liquid AI launches LFM3: six-model family with first fully omni-modal model — avapamini · 2026-10-28
• Liquid AI launches LFM3 family: 6 compact models plus d1 decision models in a single pass — antoine_chaffin · 2026-10-28
• Liquid AI releases LFM3: text, vision, audio, omni and decision models on one ~2B backbone — maximelabonne · 2026-10-28
• Liquid AI ships LFM3 small-model family; LFM3-VL-2B runs in ~1GB with big benchmark gains — mattturck · 2026-10-28

3 near-duplicate retellings: FinancialAd1961 · JosephJacks_ · maximelabonne