Liquid AI Launches LFM3: 1.7B Multimodal Models with 256K Context
By Ali Al-Montaser — Senior AI Technical Editor
Liquid AI has released the LFM3 family of six open-weight 2B multimodal models built on a shared LFM3-1.7B-Base architecture, featuring a 256K context window and 35% parameter savings. The text model outperforms LFM2.5-2.6B by a wide margin (38.42 vs. 3.02 on WebArena-Verified) and runs on a single device with 1GB memory at 4-bit precision. All models are available today on Hugging Face and customizable on a single GPU.
Executive Overview
Liquid AI has officially announced the third generation of its foundational models, LFM3, via a blog post on Hugging Face. The family comprises six multimodal 2B models built on a single shared architecture, LFM3-1.7B-Base. The models are available today for free on Hugging Face, featuring a 256K token context window, 35% parameter savings, and performance that surpasses models six times larger. This release marks a significant advancement in efficient, open-weight AI models designed for edge deployment and customization on a single GPU.
📊 Official Technical Specifications & Data Sheet
| Technical Aspect | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Free and open-weight on Hugging Face (LFM3-1.7B-Base and 5 derivative models); no licensing fees. Operational cost is limited to hosting a single GPU for customization. |
| 🌐 Platforms & Immediate Availability | Hugging Face (LFM3 collection); transformers library (5.19+ / 5.20.2+ / 5.21.0+); llama.cpp; TensorRT FP16 on Jetson AGX Orin; mobile devices, Raspberry Pi 5, PCs, and Qualcomm Hexagon NPU. |
| ⚡ Performance & Speed Benchmarks | 35% parameter savings; 25% higher speed at 8K context; WebArena-Verified: 38.42 vs. 3.02; average of 62 vision benchmarks: 71.4; UTMOSv2: 3.83; first audio token latency 50ms; LiquidCodec at 87x real-time; VLA success 96.2% on LIBERO and 45.7% on RoboCasa-GR1. |
| 🛡️ Security & Robustness | Head-wise gated attention gate curbs outlier activations affecting quantization; Sandwich norm normalization for activation stability; no explicit security certifications announced. |
| 🧠 Context Window | 256K tokens in LFM3-1.7B (after intermediate training on 3.5T tokens); SWA window of 256 tokens; 1GB memory at 4-bit precision. |
| 🌍 Arabic & Regional Support | 128K token vocabulary shared with LFM2.5-8B-A1B; 18 languages supported in audio model; no explicit Arabic support announced, but customization on a single GPU is available for Arabic data. |
Deep-Dive Features & Architecture
LFM3 is built on a Tri-hybrid architecture consisting of 24 layers: 13 gated short convolution layers, 5 sliding window attention (SWA) layers with a 256-token window, and 6 global attention layers. The convolution kernel was increased from 3 to 7, and a Head-wise gated attention gate with a sigmoid factor per head was added. Partial RoPE relaxes only the higher-frequency half in global attention layers, with Sandwich norm normalization applied to each block. As a result, LFM3-1.7B outperforms LFM2.5-2.6B with 35% parameter savings and approximately 25% higher speed at 8K context, after pre-training on 22 trillion tokens.
The shared architecture branches into six models: LFM3-1.7B for text, LFM3-VL-2B for images and video, LFM3-Audio-2B for audio and speech, LFM3-Omni-2B for all modalities, and d1-2B and d1-omni-2B for calibrated decisions. An internal vision encoder with 164 million parameters uses 3x less compute than SigLIP2 and matches Qwen3.5-2B accuracy with 2.2x fewer FLOPs. A fully causal audio encoder achieves a first-token latency of 50ms, and LiquidCodec converts tokens to speech at 87x real-time on a single CPU core of an Apple M4 Max.
Benchmark & Competitive Performance
On the WebArena-Verified agent benchmark, LFM3-1.7B scored 38.42 versus 3.02 for its predecessor LFM2.5-2.6B, surpassing Qwen3.5-9B in long-context tasks on browsers and business applications. In vision, LFM3-VL-2B raised the average across 62 benchmarks from 57.6 to 71.4 with only 1.9B parameters versus 3.1B for its predecessor, matching the six-times-larger Gemma 4 12B. In audio, it achieved a UTMOSv2 score of 3.83 versus 3.22. In robotics, a VLA trained from LFM3-VL-2B achieved 96.2% success on LIBERO and 45.7% on RoboCasa-GR1, generating motion packets in under 50ms on Jetson AGX Orin. Additionally, LFM3-Omni-2B outperforms Qwen2.5-Omni-7B on Daily-Omni, WorldSense, and AV-SpeakerBench, and d1-2B scores 75.5 versus 64.0 on Im...
Industry Impact & Enterprise Adoption
The release of LFM3 signals a major shift toward highly efficient, open-weight multimodal models that can run on edge devices without compromising performance. With a 256K context window and 1GB memory footprint at 4-bit precision, LFM3 enables advanced AI capabilities on smartphones, Raspberry Pi 5, and Qualcomm Hexagon NPUs. This democratizes access to state-of-the-art AI for enterprises and developers, allowing customization on a single GPU. The models' strong performance across text, vision, audio, and robotics benchmarks positions Liquid AI as a formidable competitor in the small-model space, challenging larger models from Qwen and Gemma. The availability of all six models on Hugging Face under a free license is expected to accelerate adoption in research and commercial applications, particularly in resource-constrained environments.
Conclusion
Liquid AI's LFM3 family represents a significant leap in efficient multimodal AI, delivering high performance with minimal computational overhead. The combination of a 256K context window, 35% parameter savings, and top-tier benchmark results makes LFM3 a compelling choice for developers and enterprises seeking powerful, customizable models. With immediate availability on Hugging Face and support for a wide range of platforms, LFM3 is poised to drive innovation in edge AI and beyond.
Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
All six LFM3 models are open-weight and available for free download on Hugging Face, including LFM3-1.7B-Base. There are no licensing fees for the models; the only cost is for hosting or a GPU for customization.
LFM3-1.7B-Base uses a 128K token vocabulary shared with LFM2.5-8B-A1B and was trained on 22 trillion tokens. The audio models support 18 languages. The official announcement does not explicitly mention Arabic support, but the model can be customized on a single GPU for Arabic data via SFT or CPT.
All LFM3 models are available today on Hugging Face within the LFM3 collection. They work via the transformers library (version 5.19 or later depending on the model) and also support llama.cpp and TensorRT FP16 on Jetson AGX Orin.
The context window in LFM3-1.7B is approximately 256K tokens, following an intermediate training phase on an additional 3.5 trillion tokens, compared to 128K in previous versions.
LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters and scores 38.42 vs. 3.02 on WebArena-Verified, surpassing Qwen3.5-9B in long-context tasks. LFM3-VL-2B also outperforms Gemma 4 12B on 62 vision benchmarks with an average of 71.4 vs. 57.6 for its predecessor.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
