Introducing LFM3: Every Modality for Specialized Intelligence

What it claims

LFM3 is Liquid AI's second-generation hybrid backbone, released as six open-weight models of about 2B parameters. On-device, the text model LFM3-1.7B decodes 1.8 to 2.8 times faster than the next-fastest compared model (llama.cpp, Q4_0, 32K context, three devices) and uses the least memory. At 128K tokens its KV cache is 1.6 GB, against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B. Liquid claims it beats Qwen3.5-9B on WebArena-Verified and AutomationBench. It also leads on BFCL and ToolSandbox. Larger models still win on agentic coding and knowledge.

How it works

The architecture came from Liquid's hardware-in-the-loop search, run this time over 14 device and backend combinations: Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, M4 Max, AMD Ryzen AI, Arm and x86 cloud CPUs, GPUs and the Qualcomm Hexagon NPU. The new pieces against LFM2:

• Sliding-window attention (256-token window) replaces some convolutions. It gave the best quality and speed tradeoff of the blocks tried.
• Convolution kernel size is 7, up from 3.
• Head-wise sigmoid gates let attention heads switch off. Liquid says this suppresses activation outliers that hurt quantization.
• Partial RoPE: global layers rotate only the high-frequency half of each head. SWA layers keep full RoPE.
• Sandwich norm normalizes each block's input and output.
• The 128K vocabulary is shared with LFM2.5-8B-A1B.

Of the 22T pre-training tokens, the final 3.5T are mid-training on curated data with more agentic traces and tool calls. This stage extends context to 256K. Post-training for the text model has four stages:

1. SFT on 1.75T tokens.
2. Mixed-domain RL with a length penalty.
3. On-policy distillation from six specialist RL teachers.
4. Agentic RL inside real harnesses.

flowchart TD
  B[LFM3-1.7B-Base] --> T[LFM3-1.7B text]
  B --> V[LFM3-VL-2B]
  B --> A[LFM3-Audio-2B]
  B --> O[LFM3-Omni-2B]
  V --> D1[d1-2B]
  O --> D2[d1-Omni-2B]

LFM3-VL uses a new 164M-parameter convolution and attention vision encoder. It uses 3× less compute than the SigLIP2 encoder in earlier LFM VL models. For video, it embeds each pair of frames into one temporal slice. The audio encoder is fully causal, so prefill runs while the user is still speaking. LiquidCodec packs 80 ms of speech into 12 tokens, which are predicted in parallel. The d1 decision models answer many typed questions in one forward pass, each with a calibrated probability. Answer types are yes/no, choice, score, Point (boxes and counts), Read (extracted span or "not found"), Estimate (value with range) and Pick (multi-select).

Evidence

• Speed: with 8K tokens in context, LFM3-1.7B is about 25% faster than LFM2-1.2B on edge devices. On short prompts it matches LFM2-1.2B per parameter.
• Throughput: on one B300 with SGLang in BF16, LFM3-1.7B produces 3.6B tokens per day at 1,024 concurrent requests, and LFM3-VL-2B exceeds 12,000 tokens per second.
• Vision: across 62 vision benchmarks (unweighted mean), LFM3-VL-2B leads same-size models by at least 7.4 points. Liquid says it matches Gemma 4 12B and is within 2 points of GPT-5.6 Luna.
• Vision encoder: it matches Qwen3.5-2B's encoder at 55.9% across 13 benchmarks with 2.2× fewer FLOPs.
• Audio: time to first token is as low as 50 ms. LiquidCodec decodes at about 87× real time on one M4 Max CPU thread. The model covers 18 languages.
• d1-Omni-2B: 84.5% on ESC-50 and 92.5% language ID over 43 languages on FLEURS.
• Wind-turbine fine-tune: about 3,000 WTBs2025 photos, SFT plus GRPO, a few hours on one A100. On about 1,400 held-out photos, classification accuracy rose to 67% from under 10%. Detection F1 rose from 0.003 to 0.11, against 0.01 for GPT-6 Luna.
• Robotics orchestration: a rank-32 LoRA on about 40,000 samples, trained in 6 GPU-hours on one MI325X, with Isaac Sim tasks.
• Robotics navigation: 2.8M samples and 72 GPU-hours give 53% on R2R and 43% on RxR, with strict stopping and no wall sliding. It runs above 2.5 Hz on a Jetson AGX Orin.
• Robotics VLA: 170 GPU-hours on 15 embodiments give 96.2% average success on LIBERO and 45.7% on RoboCasa-GR1. It produces action chunks in under 50 ms on Orin with TensorRT FP16.

How to use it

Base and post-trained weights are on Hugging Face, the Liquid Playground and the Liquid AI API. Day-0 support covers llama.cpp, MLX, vLLM and SGLang, plus OpenVINO, Qualcomm AI Hub, OpenRouter and Ollama. The text model is post-trained for the Hermes Agent, Pi and OpenClaw harnesses. Liquid says a single-GPU fine-tune can adapt it, and links fine-tuning notebooks.

Limits

The page states no licence name, so check the model card's terms before adopting. Many comparison charts are not reproduced in text, including ASR WER, TTS scores and the d1-2B versus d1-3B results. The P95 latency claims compare Liquid's proprietary RIPL server with vLLM, not like-for-like serving. The robotics manipulation and orchestration results are research extensions, mostly in simulation.