Liquid AI, the MIT spin-off behind the Liquid Foundation Models line, announced LFM3 on Tuesday, a new generation of open-weight models built to bring what the company calls "specialized intelligence" to phones, laptops, robots, and other edge hardware. The release covers six models in the 2B-parameter range — text, vision, audio, omni, and two decision models — all branched from a shared LFM3-1.7B-Base backbone and available today on Hugging Face, the Liquid AI Playground, and the company's API.
A Tri-Hybrid Architecture Redesigned for Agents For more context on this story, see our ongoing AI industry coverage.
LFM2 established that a hybrid of short convolutions and attention could run fast on-device without giving up much quality. For LFM3, Liquid AI reran its automated hardware-in-the-loop architecture search with agentic workloads in mind, expanding the search space from CPUs in one phone and one laptop to 14 device and backend combinations. The measured hardware now includes the Samsung Galaxy S26 Ultra, the OnePlus 15, a Raspberry Pi 5, Apple's M4 Max, AMD Ryzen AI laptops, Arm and x86 cloud CPUs, GPUs, and Qualcomm's Hexagon NPU.
The result is a 24-layer "tri-hybrid" stack: 13 gated short-convolution layers, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. Convolution kernel sizes grew from 3 to 7, and new mechanisms include head-wise gated attention (a sigmoid gate that can switch off individual heads), partial RoPE on global layers, and sandwich normalization. The company says the changes add expressivity where it counts while keeping long-context inference fast.
Trained on 22 Trillion Tokens
LFM3-1.7B-Base was pre-trained on 22 trillion tokens, with a data mix that adds more scientific papers, repository-level code, and math than LFM2. The final 3.5 trillion tokens form a mid-training stage on curated, higher-quality data that is heavy on agentic traces and diverse tool calls — and that stage also extends the context window to 256K tokens.
Post-training ran through four stages: supervised fine-tuning on 1.75 trillion tokens spanning chat, reasoning, and tool calling; mixed reinforcement learning across math, code, SQL, and structured outputs; multi-domain on-policy distillation, in which six specialist RL teachers (code, math, tool use, long context, agentic work, and enterprise workflows) are distilled back into one model; and agentic RL, where tasks from web browsing to office work run inside real agent harnesses such as Hermes Agent, Pi, and OpenClaw.
Small Model, Frontier-Adjacent Agentic Results
Liquid AI claims LFM3-1.7B outperforms its previous LFM2.5-2.6B while using 35 percent fewer parameters. On agentic evaluations, the company says the model leads on instruction following and function calling in BFCL and ToolSandbox, and beats the substantially larger Qwen3.5-9B on long-horizon browser and business-app tasks in WebArena-Verified and AutomationBench. Larger models still win on agentic coding and knowledge, though the company reports that tool search nearly doubles the small model's score on Humanity's Last Exam.
Speed is the headline. Measured with llama.cpp at 4-bit quantization and a 32K context, Liquid AI says LFM3-1.7B was the fastest model on all three tested edge devices, decoding 1.8 to 2.8 times faster than the next competitor while using the least memory. A full 128K-token context needs about 1.6 GB of KV cache, versus 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B, and the model fits in roughly 1 GB of memory at 4-bit precision. In the datacenter, a single NVIDIA B300 running SGLang in BF16 can reportedly produce 3.6 billion tokens per day at 1,024 concurrent requests. To show what that throughput enables, Liquid AI ran more than 300,000 parallel LFM3-1.7B agents over the Enron email archive as a demo.
Vision, Audio, and a First Omni Model
LFM3-VL-2B brings a new in-house vision encoder with 164 million parameters that, according to Liquid AI, uses 3x less compute than the SigLIP2 encoder in previous LFM vision models while matching Qwen3.5-2B's encoder accuracy with 2.2x fewer FLOPs. Across 62 vision benchmarks, the company says the 2B model scores at least 7.4 points ahead of all similar-size models, matches Gemma 4 12B — a model more than six times larger — surpasses Gemini 3.5 Flash Lite, and lands within 2 points of GPT-5.6 Luna. It is also the first LFM vision model to accept video input.
LFM3-Audio-2B handles speech recognition, synthesis, and translation across 18 languages, powered by a new in-house encoder and a tokenizer called LiquidCodec that converts generated tokens to speech at roughly 87x real time on a single CPU thread. With streaming microphone input, time to first token can be as low as 50 milliseconds.
LFM3-Omni-2B, at 2.2 billion parameters, is the company's first omni model, combining vision, audio, and language so users can ask spoken questions about images or about videos with synchronized sound. Two decision models, d1-2B and d1-Omni-2B, close out the family: instead of generating free-form answers, they answer many typed or spoken questions in a single forward pass, each with a calibrated probability, and support structured answer types like Point (draw a bounding box), Read, Estimate, and Pick.
Built to Be Specialized
Liquid AI positions the family as a starting point for fine-tuning rather than a general-purpose frontier model. In one demonstration, the company fine-tuned LFM3-VL-2B on about 3,000 public photos of wind turbine blades to classify defects. Out of the box, both LFM3-VL-2B and GPT-6 Luna correctly identified the defect types in fewer than 10 percent of test images; after a few hours of training on a single A100, the fine-tuned LFM3 reached 67 percent classification accuracy and improved detection F1 from 0.003 to 0.11, compared with 0.01 for GPT-6 Luna. The company also published early robotics experiments covering tool orchestration, vision-language navigation, and low-level robot-arm manipulation.
Day-zero launch partners include Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm via AI Hub and GenieX, and Vercel, with Mercedes-Benz and Shopify participating in pre-release testing. Community runtimes including llama.cpp, MLX, LM Studio, Ollama, and vLLM support the models at launch.
The release underlines a broader industry split: as frontier labs chase ever-larger generalist models, Liquid AI is betting that small, fast, specialized models — cheap enough to fine-tune on one GPU and small enough to run locally — will be where much of the practical value lands.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
