LFM3 Review: Liquid AI Open-Weight 2B Models for Local Deployment

LFM3 review examines the six models Liquid AI released on October 27, 2026. Each shares the 1.73-billion-parameter LFM3-1.7B-Base backbone trained on 22 trillion tokens with a tri-hybrid architecture of 24 layers.

The models target specialized on-device workloads rather than general frontier tasks. All weights are available on Hugging Face under the LFM Open License v1.0.

LFM3 Architecture and Modalities

The base model uses 13 gated short convolutions, five sliding-window attention layers and six global attention layers with a 128K vocabulary. Post-trained variants extend context and add modalities while keeping parameter counts near 2 billion.

Six released models cover different input and output combinations:

• LFM3-1.7B: text in and out, 256K context, focused on agents and tool calling.
• LFM3-VL-2B: text, images and video in, text out, 128K context.
• LFM3-Audio-2B: text and audio in, text and speech out, 32K context across 18 languages.
• LFM3-Omni-2B: text, images, video and audio in, text and speech out, 32K context.
• d1-2B and d1-Omni-2B: multimodal inputs with calibrated single-pass decision outputs.

Weights appear in GGUF, MLX, ONNX, FP8 and NVFP4 formats. The license permits free commercial use for companies under $10 million annual revenue and unrestricted research use.

Local Deployment, Pricing and Performance

Quantized to 4 bits, the text model occupies roughly 1 GB and runs on phones, laptops and edge devices including iPhone 17 Pro, Apple M4/M5, AMD Ryzen AI and Raspberry Pi 5. Primary runtimes include llama.cpp, MLX, vLLM and partial Ollama support.

API pricing through Liquid Console and OpenRouter lists LFM3-1.7B at $0.08 per million input tokens, $0.02 cached and $0.20 output. Multimodal variants carry higher rates between $0.10 input and $0.50–1.00 output per million tokens.

Vendor benchmarks report LFM3-1.7B averaging 49.37 across 29 tests, exceeding Qwen3.5-9B on the same suite. Strong results appear on WebArena-Verified (38.42) and BFCL V4 tool calling (71.39). Independent local tests confirm reliable structured output yet note weaknesses in complex coding and some multilingual tasks.

Limitations and Comparison to Alternatives

Model cards advise against using LFM3 for complex agentic coding or multi-step reasoning without retrieval. Video support remains limited outside Liquid frameworks, and full Ollama integration is still partial.

Compared with Qwen3.5-1.5B and Phi-3-mini, LFM3 offers longer context on the text model and native video in the VL variant. Both alternatives provide fully permissive licenses without revenue thresholds. LFM3 edges ahead in reported agent and tool-calling scores on Liquid’s internal suite but trails larger models on knowledge-intensive benchmarks.

Users seeking unrestricted commercial redistribution or broader runtime support may prefer fully open alternatives, while teams prioritizing on-device speed and calibrated decision outputs find the current LFM3 release practical for targeted workloads.