New LFM3 Family Puts Five Modalities on One 2B-Scale Backbone
LFM3 puts text, vision, audio, omni and decision models on one 2B-scale backbone for single-GPU tuning and edge use.
The post listed four main changes from the LFM2 family: more capability per parameter, every modality on one backbone, edge-oriented design and support for specialization. It said LFM3-1.7B outperforms LFM2.5-2.6B while using 35% fewer parameters.
The base model has 24 layers and uses a tri-hybrid layer mix: 13 gated short convolutions, five sliding-window attention layers with a 256-token window and six global attention layers. The convolutions use a kernel size of 7, up from 3 in LFM2. LFM3 adds head-wise gated attention, in which a sigmoid gate scales each head's output so a head can switch off when it has nothing to add; partial RoPE, which rotates only the highest-frequency half of each head in global attention layers; and sandwich norm, which normalizes each block's input and output.
The post said hardware-in-the-loop architecture search covered 14 device and backend combinations, from phones and a Raspberry Pi 5 to laptops, cloud CPUs, GPUs and the Qualcomm Hexagon NPU. LFM3-1.7B-Base was pre-trained on 22T tokens with the same 128K vocabulary as LFM2.5-8B-A1B. The final 3.5T tokens were a mid-training stage on curated data with a larger share of agentic traces and tool calls, which also extended the context window to 256K tokens.
The text model fits in about 1 GB at 4-bit precision, and the 256K context window is intended to let agents run on-device. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context, according to the post.
Every LFM3 model branches from LFM3-1.7B-Base. The lineup includes LFM3-1.7B for text-to-text; LFM3-VL-2B for text, images and video to text; LFM3-Audio-2B for text and audio to text and speech; LFM3-Omni-2B for text, images, video and audio to text and speech; d1-2B for text, images and video to calibrated answers; and d1-omni-2B for text, images, video and audio to calibrated answers. The post described LFM3-Omni-2B as its first omni model and said the vision model adds video input.
The vision and audio models use new in-house encoders. The vision encoder combines convolution and attention and uses three times less compute than the SigLIP2 encoder in previous LFM VL models, according to the post. A 164M-parameter encoder matches Qwen3.5-2B's vision encoder in accuracy while requiring 2.2 times fewer FLOPs, measured with the same LFM2.5-230M language backbone. For video, it embeds each pair of sampled frames into one temporal slice, halving the temporal sequence length.
The audio encoder mixes sliding-window attention and short convolutions and is fully causal, so the model can prefill while a user speaks, with time to first token as low as 50 ms. LiquidCodec, the new audio tokenizer, turns generated tokens into speech at about 87 times real time on a single CPU thread on Apple M4 Max and supports streaming decoding. The post said each model runs in transformers.
LFM3-1.7B is positioned for local agents, tool calling, structured outputs and domain-specific fine-tuning. It is post-trained to run in agentic harnesses such as Hermes Agent, Pi and OpenClaw, according to the post. On WebArena-Verified, it scores 38.42 against 3.02 for its predecessor, and it beats Qwen3.5-9B on long tasks in browsers and business apps, the post said.