Latest: Liquid AI's LFM3 Six-Model Family Lands on NVIDIA RTX Spark
The family's efficiency claims, if independently verified, would represent meaningful on-device capability at sub-2B parameter scale. All benchmark scores remain vendor-reported with no third-party reproduction, so the performance claims carry that caveat.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
Liquid AI released LFM3 on October 27, 2026, as six open-weight models covering text, vision-language, audio, omni, and decision-making tasks. The family shares a tri-hybrid architecture combining short-convolution, sliding-window attention, and global attention layers, optimized across 14 processor classes for on-device inference and single-GPU fine-tuning. NVIDIA has since added LFM3 to its RTX Spark laptop platform, where models run locally via llama.cpp and vLLM and support agent workflows including inbox triage, calendar management, document reading, and speech transcription.
Liquid AI reports the 1.7B text model scores 41.29 on AutomationBench versus 15.10 for Qwen3.5-9B, and outperforms the company's prior 2.6B model. LFM3-VL-2B, at 1.9B parameters, scores at least 7.4 points ahead of all models in its size class across 62 vision benchmarks per Liquid AI, matches Gemma 4 12B, surpasses Gemini 3.5 Flash Lite, and comes within 2 points of GPT-5.6 Luna; Liquid AI also states it achieves the highest prefill throughput and lowest memory footprint in its size class on-device. LFM3-Omni-2B is a 2.17B dense-parameter model accepting text, image, video, and audio inputs and producing text and 24 kHz speech output with a 32,768-token context window; Liquid AI claims it beats models up to 10.7B parameters, including Qwen2.5-Omni-7B, on Daily-Omni, WorldSense, and AV-SpeakerBench. All benchmark figures across the family are vendor-reported by Liquid AI.
NVIDIA added LFM3 to its RTX Spark laptop platform, where the six open-weight models Liquid AI released on October 27 run locally via llama.cpp and vLLM and support agent workflows including inbox triage, calendar management, document reading, and speech transcription.
Developers can also fine-tune LFM3 models locally on RTX Spark hardware.
Liquid AI released LFM3 on October 27, 2026, six open-weight models for text, vision-language, audio, omni, and decision tasks on a tri-hybrid architecture built for on-device inference.
LFM3-VL-2B claims at least a 7.4-point lead across 62 vision benchmarks in its size class and matches Gemma 4 12B, while LFM3-Omni-2B, at 2.17B parameters with a 32,768-token context window, claims to beat models up to 10.7B parameters on three benchmarks. Every benchmark figure is vendor-reported by Liquid AI with no published methodology and no third-party reproduction.
BackgroundFrontier models, open models, and how they are compared
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.