01 / The short version
What happened
LFM3 introduces a family of specialized multimodal foundation models in the 2B parameter range, built on a shared backbone that unifies text, vision, audio, omni, and decision capabilities. The architecture was redesigned for agentic use cases using hardware-in-the-loop search across 14 device and backend combinations. LFM3-1.7B matches LFM2-1.2B on short prompts and is 25% faster at 8K context. LFM3-VL-2B raises average scores on 62 vision benchmarks to 71.4, on par with Gemma 4 12B. LFM3-Audio-2B improves UTMOSv2 to 3.83. LFM3-Omni-2B beats models up to 10.7B parameters on audio-visual benchmarks. d1 models provide structured decision outputs in a single forward pass.
See the exact references02 / Key takeaways
What you need to know
- 01
LFM3 unifies text, vision, audio, omni, and decision models in 2B parameter range
- 02
LFM3-VL-2B achieves 71.4 average on 62 vision benchmarks, matching Gemma 4 12B
- 03
d1 models deliver structured decisions in a single forward pass with calibrated probabilities
Keep in perspective
What to watch for
Performance claims are attributed to the publisher; independent validation of benchmark results is not provided.
Go to the source
Exact references
These are the original pages used for this brief. Publisher claims are not independent evaluations.
01Primary source · Hugging FaceRead the original announcementhttps://huggingface.co/blog/LiquidAI/lfm3AI-generated from the linked source. It can miss context; verify consequential details in the original. How the radar works