Liquid AI Unveils LFM3: Six Models, 256K Context, and a New Tri-Hybrid Architecture

The lineup features a new tri-hybrid architecture, support for 14 processor classes, and single-GPU fine-tuning.

The company claims its 1.9B-parameter LFM3-VL-2B competes with Gemma 4 12B, Gemini 3.5 Flash Lite, and GPT-5.6 Luna.

Six Models Across Four Modalities

The d1 models generate typed decisions with calibrated probabilities in a single inference pass rather than relying on free-form text generation.

LFM3 vs. LFM2.5: 8× More Context, Fewer Vision Parameters

LFM3 increases text context capacity by 8× while reducing the flagship vision model's parameter count by approximately 39%.

These figures describe architectural and capacity changes, not measured improvements in accuracy or speed.

Tri-Hybrid Architecture Targets 14 Processor Classes

LFM3 replaces the previous hybrid convolution-and-attention design with a new tri-hybrid architecture.

Liquid AI reports three deployment capabilities:

• Optimization for 14 processor classes across CPUs, GPUs, and NPUs.
• Single-GPU fine-tuning.
• On-device inference across text, vision, audio, and decision models.

Detailed architectural components, GPU memory requirements, and fine-tuning configurations were not provided in the announcement.

A 1.9B Vision Model Competing With 12B-Class Models

Liquid AI claims LFM3-VL-2B delivers competitive performance against significantly larger models.

For reference, the previous LFM2.5-VL-3B achieved the following company-reported results:

These are previous-generation benchmark results, not LFM3 scores. The LFM3 announcement does not include numerical results or standardized testing configurations to independently validate its competitive claims.

LFM3-1.7B Tops Liquid AI's Device-Efficiency Pareto Frontier

Liquid AI positions LFM3-1.7B at the top of the Pareto frontier in its Pipette device-intelligence benchmark.

Pipette evaluates model quality and deployment efficiency, including inference throughput, latency, and memory requirements.

However, the announcement does not disclose exact LFM3 measurements for tokens per second, time to first token, VRAM consumption, or energy efficiency.

Where the Six Models Fit

What Remains Unverified

LFM3 brings an 8× larger text context window, a smaller vision-language model, integrated multimodal processing, and new decision-focused variants.

Its claimed performance against Gemma 4 12B and other competitors remains unverified without published LFM3 benchmark scores, latency measurements, and hardware configurations.