Liquid AI LFM3 2B-parameter multimodal models released

LFM3 delivers multimodal, on‑device intelligence in a 2 B‑parameter family

Takeaway: Liquid AI released LFM3, a 2 B‑parameter multimodal model family (text, vision, audio, omni, and decision) that runs on phones, laptops, and edge CPUs/GPUs, and beats larger models on a wide range of agentic and multimodal benchmarks.

New hybrid architecture gives more expressivity per parameter

  • LFM3 uses a tri‑hybrid layer mix: 13 gated short‑convolution layers, 5 sliding‑window attention (SWA) layers with a 256‑token window, and 6 global‑attention layers across 24 total layers. Convolution kernels are size 7 (up from 3 in LFM2).
  • Head‑wise gated attention applies a sigmoid gate to each attention head, allowing heads to deactivate when they add no value, which improves quantization stability.
  • Partial RoPE rotates only the high‑frequency half of each head in global‑attention layers, preserving low‑frequency content for distance‑invariant matching.
  • Sandwich norm normalizes both block input and output, keeping activations well‑scaled.
  • The model is trained on 22 T tokens, with the final 3.5 T tokens emphasizing higher‑quality, agentic traces and tool‑call data, and extending the context window to 256 K tokens.
  • Hardware‑in‑the‑loop architecture search was expanded from 2 devices (LFM2) to 14 device/backend combos, including Samsung Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI, Arm and x86 CPUs, GPUs, and Qualcomm Hexagon NPU.

Result: Per‑parameter, LFM3‑1.7B matches LFM2‑1.2B on short prompts and is ~25 % faster than LFM2 when processing 8 K‑token contexts.

Multimodal family covers every modality

Model Modality Parameters Key capability
LFM3‑1.7B Text 1.7 B 256 K‑token context, agentic post‑training, tool use
LFM3‑VL‑2B Vision (image + video) ~2 B Video Q&A, step‑by‑step reasoning, spatial grounding
LFM3‑Audio‑2B Audio ~2 B Speech‑to‑text, multilingual TTS, audio reasoning, tool calling
LFM3‑Omni‑2B Vision + Audio + Text 2.2 B Spoken visual QA, audio‑visual video understanding, multimodal tool calls
d1‑2B Decision (text + image) 2 B Single‑pass structured answers with calibrated probabilities
d1‑Omni‑2B Decision (text + image + audio) 2 B Same as d1‑2B but accepts spoken queries

All models share the same backbone, allowing developers to pick the smallest model that fits their use case and fine‑tune it locally.

Text model LFM3‑1.7B excels on agentic tasks

  • Benchmark highlights (higher is better unless noted):
    • Instruction following (IFEval): 96.66 % – top of the table.
    • Tool use (ToolSandbox): 80.36 % – beats Qwen3.5‑9B.
    • Long‑context tasks (RULER 128K): 65.56 % – outperforms larger 8‑12 B models.
    • Agentic coding (SWE‑bench Verified): 43.40 % – comparable to Qwen3.5‑4B.
  • Throughput: On an NVIDIA B300 with SGLang BF16, LFM3‑1.7B generates 3.6 B tokens/day at 1 024 concurrent requests, the highest among the compared models.
  • On‑device speed: Measured with llama.cpp Q4_0, LFM3‑1.7B decodes 1.8–2.8× faster than the next fastest model and uses the least memory; a 128 K‑token KV cache occupies 1.6 GB, versus 4.3 GB for Qwen3.5‑4B.
  • Real‑world agent test: In a Mac‑based evaluation of 65 everyday tasks, LFM3‑1.7B achieved 64 % task success (vs 48 % for Qwen3.5‑2B), used 58 % fewer tokens, and stalled on <1 % of attempts.

Vision model LFM3‑VL‑2B sets new efficiency standards

  • Performance: Across 62 vision benchmarks, LFM3‑VL‑2B scores an average of 71.4, beating all same‑size competitors by at least 7.4 points and matching the 12 B‑parameter Gemma 4 on many tasks.
  • Video understanding: First LFM vision model that accepts video input; can answer questions about lectures, walkthroughs, or camera feeds.
  • Efficiency: New in‑house vision encoder (164 M parameters) matches Qwen3.5‑2B’s accuracy while using 2.2× fewer FLOPs. On‑device prefill throughput is the highest in its class.
  • Throughput: On an NVIDIA B300 with SGLang BF16, LFM3‑VL‑2B produces >12 k tokens/s, surpassing all comparable models.

Audio model LFM3‑Audio‑2B raises the bar for on‑device speech

  • Quality metrics:
    • UTMOSv2 (speech naturalness): 3.8 – higher than Qwen2.5‑Omni‑3B (3.4).
    • WER (English ASR): 1.4 % – best among listed models.
    • VoiceBench: 73.1 – top score, indicating strong voice generation and tool‑calling.
  • Architecture: Combines sliding‑window attention and short convolutions in a causal audio encoder; uses LiquidCodec tokenizer that compresses 80 ms of speech into 12 discrete tokens, enabling parallel token prediction.
  • Latency: First token appears within ≈50 ms after speech ends, supporting natural‑language voice assistants.

Omni model LFM3‑Omni‑2B unifies vision, audio, and language

  • Cross‑modal capabilities: Spoken visual QA, audio‑visual video reasoning, and multimodal tool calling in a single 2.2 B‑parameter model.
  • Benchmark comparison: Beats Qwen2.5‑Omni‑3B on most video+audio tasks (e.g., Daily‑Omni 70.1 % vs 58.8 %).
  • On‑device use‑case: Enables assistants that can see, hear, and act without cloud connectivity, suitable for tablets, in‑car systems, and robotics.

Decision models provide instant structured answers

  • d1‑2B (text + image) and d1‑Omni‑2B (adds audio) replace the previous d1‑3B with ~33 % fewer parameters while improving accuracy:
    • Decision Bench (enterprise records): 85.2 % vs 83.6 %.
    • Decision Index: 51.1 vs 49.0.
  • Typed answers now include Point (bounding boxes), Read (extracted spans), Estimate (numeric ranges), and Pick (multi‑option selection) in a single forward pass.
  • Spoken queries retain most of the accuracy (e.g., DecisionBench 78.2 % spoken vs 83.4 % written).

Community feedback from Hacker News

"Excited to roll out LFM 3 to replace LFM 2.6; looks like a big step up." – @colingauvin

"I worked on this release, happy to answer questions about local AI vs cloud frontier models." – @samstevens

"LFM3‑VL‑2B gave better results than LFM3‑1.7B on text‑only prompts; likely a quantization issue in the 1.7B Q8 version." – @lubitelpospat

"Agentic coding numbers on par with Qwen3.5‑4B; impressive for a 1.7B model!" – @jauntywundrkind

These comments highlight early adopters’ enthusiasm, the practical advantage of on‑device deployment, and a few quantization quirks that users are already investigating.

Getting started

  • Weights: Open‑weight checkpoints are available on Hugging Face, Liquid AI Playground, and via the Liquid AI API.
  • Supported hardware: Apple M4 Max, AMD Ryzen AI, Qualcomm Hexagon NPU, Arm/x86 CPUs, NVIDIA GPUs, and cloud back‑ends through SGLang, llama.cpp, MLX, and vLLM.
  • Fine‑tuning: The 2 B‑class size enables full fine‑tuning on a single GPU; the blog provides notebooks for domain‑specific adaptation (e.g., wind‑turbine defect detection) and robotics pipelines.

Why LFM3 matters

LFM3 proves that a 2 B‑parameter multimodal backbone can outperform much larger models on specialized, agentic, and long‑context tasks while running efficiently on edge devices. This shifts the frontier from cloud‑only, high‑cost inference toward private, low‑latency, on‑device AI that can be customized by developers without massive compute budgets.

Sources