Liquid AI releases LFM3 family of six open-weight 2B-class multimodal models

On October 27, 2026, Liquid AI introduced LFM3, a family of six open-weight 2B-class multimodal models spanning text, vision, audio, omni, and decision tasks. Built on a tri-hybrid architecture combining convolutions and attention mechanisms across 24 layers, the flagship LFM3-1.7B fits into roughly 1GB at 4-bit quantization and outperforms Qwen3.5-9B on web agent and tool-use benchmarks like WebArena-Verified (38.42 vs 28.94), though it trails on complex coding and general knowledge tasks.

LFM3 family release and model variants

• ▪Liquid AI released LFM3 under the LFM Open License v1.0, requiring commercial organizations generating over $10 million in annual revenue to obtain a separate commercial agreement
• ▪Liquid AI identified Mercedes-Benz and Shopify as testing partners evaluating the LFM3 model family prior to its October 27, 2026 release
• ▪The Liquid AI LFM3 family includes LFM3-1.7B for text agents, LFM3-VL-2B for vision, LFM3-Audio-2B for speech, LFM3-Omni-2B for multimodal input, and the d1-2B and d1-Omni-2B decision models
• ▪Liquid AI released LFM3, a family of six open-weight 2-billion-parameter-class multimodal AI models built on a shared base backbone

Architecture and training pipeline

• ▪Liquid AI's LFM3-1.7B base model was pretrained on 22 trillion tokens, with the final 3.5 trillion tokens dedicated to agentic mid-training to extend its context window to 256,000 tokens
• ▪Liquid AI's LFM3 backbone features a tri-hybrid architecture combining 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers across 24 layers
• ▪Liquid AI developed the LFM3 tri-hybrid architecture using hardware-in-the-loop architecture search evaluated across 14 device and backend combinations, including the Raspberry Pi 5 and Qualcomm Hexagon NPU
• ▪Liquid AI's LFM3 post-training pipeline comprised four stages: supervised fine-tuning on 1.75 trillion tokens, mixed reinforcement learning, distillation of six specialist teachers into one model, and agentic reinforcement learning

Benchmark evaluation performance

• ▪Liquid AI's LFM3-1.7B scored 63.37 on GPQA Diamond and 43.40 on SWE-bench Verified, trailing Qwen3.5-9B's respective scores of 80.40 and 51.20
• ▪On the WebArena-Verified evaluation, Liquid AI's LFM3-1.7B scored 38.42, outperforming Qwen3.5-9B's score of 28.94 and Liquid AI's previous LFM2.5-2.6B score of 3.02
• ▪On the AutomationBench public evaluation, Liquid AI's LFM3-1.7B achieved a score of 41.29, compared to 15.10 for Qwen3.5-9B and 0.94 for LFM2.5-2.6B
• ▪On the RULER 128K context evaluation, Liquid AI's LFM3-1.7B scored 65.56, trailing Qwen3.5-9B's score of 87.11
• ▪On the BFCL V4 function-calling benchmark, Liquid AI's LFM3-1.7B scored 71.39, exceeding Qwen3.5-9B's score of 62.83 and LFM2.5-2.6B's score of 55.89

Local memory usage and footprint

• ▪Liquid AI reported that LFM3-1.7B fits into approximately 1GB of storage at 4-bit precision, enabling local deployment on laptops and mobile phones
• ▪Processing a 128,000-token context with Liquid AI's LFM3-1.7B requires 1.6GB of KV cache working memory during generation, compared to 4.3GB for Qwen3.5-4B
• ▪During local testing with a 32,000-token input, Liquid AI's LFM3-1.7B recorded peak memory usage of 2.75GB on an AMD Ryzen AI Max+ 395 system, 2.25GB on an Apple M5 Max, and 1.90GB on an iPhone 17 Pro

Formats, runtimes, and GPU throughput

• ▪Liquid AI reported that a single Nvidia B300 GPU running SGLang with LFM3-1.7B generated an aggregate throughput of 3.6 billion tokens daily at 1,024 concurrent requests
• ▪Liquid AI released LFM3 checkpoints in GGUF, MLX, ONNX, FP8, and NVFP4 formats, supporting runtimes including llama.cpp, vLLM, SGLang, Ollama, and Transformers
• ▪In latency benchmarks on a single Nvidia B200 GPU, Liquid AI recorded a 95th-percentile completion time of 150 milliseconds at 500 requests per second for short requests of 128 input and 64 output tokens

Vision model fine-tuning

• ▪Liquid AI demonstrated local customization by fine-tuning LFM3-VL-2B on 3,000 wind-turbine images over several hours on an Nvidia A100 GPU, increasing classification accuracy from below 10% to 67%
• ▪Liquid AI adapted the LFM3 vision model to operate a simulated robot arm through software tools using approximately six GPU-hours on an AMD Instinct MI325X

API pricing and privacy policy

• ▪Liquid AI states that its authenticated API console enforces zero data retention, asserting that user-submitted content is not stored, logged, or used for model training
• ▪Liquid AI offers hosted LFM3-1.7B API access via its developer console priced at $0.08 per million input tokens, $0.02 per million cached input tokens, and $0.20 per million output tokens