LFM3: New Multimodal Foundation Model for Edge AI and Agents

LFM3 represents a significant leap in the development of Liquid Foundation Models (LFMs), establishing a new standard for specialized, multimodal intelligence. Unlike previous generations, LFM3 is engineered around a single, shared backbone architecture, allowing it to seamlessly integrate and process multiple modalities—including text, vision, audio, omni-directional data, and complex decision-making processes—all within a highly efficient 2 billion parameter range. This architectural design is revolutionary because it moves away from the need for massive, monolithic general-purpose models, instead empowering developers to select, fine-tune, and deploy highly specialized models tailored precisely to their application's unique data, terminology, and operational environment. The core breakthrough lies in achieving high capability with drastically reduced parameter counts, making advanced AI accessible for deployment on resource-constrained edge devices.

Compared to its predecessor, the LFM2 family, LFM3 introduces several critical architectural and training enhancements. One of the most impressive advancements is the efficiency gain: the LFM3-1.7B model demonstrates superior performance compared to the LFM2.5-2.6B model, yet it achieves this while utilizing approximately 35% fewer parameters. This efficiency is not accidental; it is the result of a sophisticated new tri-hybrid architecture and a refined training recipe. Furthermore, LFM3 expands the scope of multimodality by offering six distinct 2B-class models that all branch from the LFM3-1.7B-Base foundation. This expanded suite includes the first dedicated omni model and crucially incorporates video input capabilities into the vision model, broadening its applicability from simple image recognition to complex temporal reasoning.

Crucially, LFM3 is built with the edge device in mind. The text-only model, for instance, can be compressed to fit into approximately 1 GB when quantized to 4-bit precision, a massive boon for local, on-device AI applications. This efficiency is supported by new in-house vision and audio encoders designed to accelerate processing. Moreover, the text model boasts an extensive 256K context window, a feature that is vital for enabling sophisticated AI agents to operate autonomously on local hardware. The entire system is designed to be modular, allowing developers to fine-tune each specialized model on a single GPU, ensuring maximum customization without requiring massive cloud infrastructure.

Under the hood, the architectural improvements are highly technical and sophisticated. The LFM3-1.7B-Base model features a novel tri-hybrid layer mix. This mix comprises 13 gated short convolutions, 5 sliding-window attention (SWA) layers that operate within a 256-token window, and 6 global attention layers. The convolutions utilize an increased kernel size of 7, up from the previous 3 used in LFM2. A key innovation is the implementation of head-wise gated attention, which uses a sigmoid gate to dynamically scale each attention head's output. This mechanism allows the model to effectively 'switch off' any head that is not contributing meaningful information, thereby suppressing activation outliers that often degrade performance during quantization. Complementing this is the use of Partial RoPE (Rotary Position Embedding) in the global attention layers, which only rotates the highest-frequency half of each head, ensuring that the remaining half remains content-agnostic regardless of the distance between tokens. Finally, the adoption of the Sandwich norm ensures that both the input and output of every block are properly normalized, maintaining stable and well-scaled activations throughout the network.

These architectural choices yield substantial performance gains, particularly in long-context scenarios. On a per-parameter basis, LFM3-1.7B can match the performance of LFM2-1.2B on short prompts across various edge devices. More dramatically, when handling a context window of 8K tokens, LFM3 proves to be approximately 25% faster. The model's training regimen is equally advanced: LFM3-1.7B-Base was initially pre-trained on a massive corpus of 22 Trillion tokens, sharing the same 128K vocabulary as LFM2.5-8B-A1B. Critically, the final 3.5 Trillion tokens of mid-training were curated specifically on data containing a larger proportion of agentic traces and tool calls. This specialized training not only extends the context window to 256K tokens but also hardwires the model's ability to perform complex, multi-step reasoning required for autonomous agents.

Instead of forcing users to run a single, massive general-purpose model, the LFM3 framework mandates a modular approach. Every specialized LFM3 model branches directly from the LFM3-1.7B-Base. This allows developers to select the exact model that matches their input type—be it text, vision, or audio—customize it for their domain, and run it efficiently on the device. The vision and audio modalities also benefit from entirely new, optimized in-house encoders. The vision encoder, for instance, combines convolution and attention and achieves a remarkable reduction in computational load, requiring three times less compute than the SigLIP2 encoder used in previous LFM vision models. While maintaining accuracy comparable to Qwen3.5-2B, it demands 2.2 times fewer Floating Point Operations (FLOPs). For video processing, the system further optimizes temporal data by embedding each pair of sampled frames into a single temporal slice, effectively halving the sequence length and improving efficiency. The audio pipeline is equally advanced, featuring an audio encoder that mixes SWA and short convolutions, making it fully causal. This causality allows the model to begin processing (prefill) while the user is still speaking, achieving an extremely low time-to-first-token of as little as 50 milliseconds. Furthermore, the accompanying LiquidCodec tokenizer converts generated tokens into speech at an astonishing rate of about 87 times real-time on a single CPU thread on Apple M4 Max, supporting seamless streaming decoding.

In terms of practical application, LFM3-1.7B is specifically recommended for use cases requiring local agents, robust tool calling capabilities, structured output generation, or as a foundational base for domain-specific fine-tuning. It is pre-trained to operate within agentic harnesses such as Hermes Agent, Pi, and OpenClaw. Its performance gains are most pronounced in agentic tasks: on the WebArena-Verified benchmark, it scores 38.42, a massive improvement over its predecessor's score of 3.02, and it even surpasses the performance of Qwen3.5-9B on complex, long-running tasks within browser and business application simulations. Similarly, the LFM3-VL-2B model, designed for on-device visual understanding, is noted as the first LFM vision model to accept video input. It is capable of step-by-step reasoning and object placement analysis, making it a powerful foundation for advanced robotics applications. Compared to LFM2.5-VL-3B, LFM3-VL-2B uses fewer parameters (1.9B vs 3.1B) yet raises the average score across 62 vision benchmarks from 57.6 to 71.4, achieving parity with the performance of the much larger Gemma 4 12B model, solidifying its position as a leading, efficient, and highly capable multimodal AI solution.