LFM3-VL Review: Liquid AI's 2B Vision-Language and Audio Models

LFM3-VL-2B and LFM3-Audio-2B form two members of Liquid AI’s six-model LFM3 family released October 27, 2026. Both models build on the shared LFM3-1.7B-Base backbone pretrained on 22 trillion tokens and target on-device multimodal workloads. The VL variant processes images and video at native resolution while the Audio variant handles 16 kHz input and produces 24 kHz speech output.

LFM3-VL Review: Vision Capabilities and Benchmarks

The LFM3-VL-2B model contains 1.92 billion total parameters: 1.73 billion in the language model, 164 million in an in-house hybrid vision encoder, and 25 million in the projector. It accepts 128K context length and supports text plus still images or video sampled at 2 frames per second with timestamps. Output remains text only. The model supports step-by-step reasoning mode and 16 languages under the LFM Open License v1.0.

Vendor benchmarks place LFM3-VL-2B at an average of 71.4 across 62 vision tasks, compared with 57.6 for the prior LFM2.5-VL-3B. Strong categories include document and chart reading at 83.9, GUI understanding at 83.1, and visual grounding at 77.3. Weaker areas appear in video tasks at 61.2 and agentic scenarios at 42.5. These figures sit close to Gemma 4 12B at 71.7 and within two points of GPT-5.6 Luna at 73.4.

The vision encoder uses three times less compute than SigLIP2 equivalents, enabling deployment on phones, laptops, and edge hardware. Fine-tuning examples already include industrial defect detection and robotics adaptation. The model is not positioned for coding or complex multi-turn reasoning.

LFM3-Audio Review: Speech Features, Pricing, and Access

LFM3-Audio-2B operates at roughly 2 billion parameters with a 32K context window. It accepts text or 16 kHz audio and returns text, 24 kHz speech, or interleaved text-plus-speech. An in-house causal audio encoder paired with LiquidCodec supports streaming, ASR, TTS, audio understanding, and spoken tool calling across 18 languages. Time-to-first-token reaches as low as 50 ms, and UTMOSv2 TTS quality improved from 3.22 to 3.83 versus the LFM2.5 predecessor.

Reported scores include VoiceBench at 73.1, English ASR word-error rate near 6.2 percent, and internal tool-calling at 58.5. Weights appear on Hugging Face under the LiquidAI organization in BF16, quantized, GGUF, MLX, and ONNX formats. Users can run the models through Transformers, llama.cpp, vLLM, SGLang, or Ollama. Liquid AI also offers a playground and API access.

API pricing stands at approximately $0.10 per million input tokens and $0.50 per million output tokens for the VL model; the Audio model lists $0.10 input and $1.00 output. Cached input receives lower rates. GGUF variants for the Audio model remain forthcoming, and independent benchmark suites are still in progress.

Comparison With Similar Open Multimodal Models

Both LFM3 models emphasize single-GPU fine-tuning and on-device optimization. They differ from larger closed models by releasing full weights and supporting multiple inference runtimes. Users seeking multilingual speech output or native video input at 2B scale currently have few direct open alternatives with matching encoder efficiency.