LFM3 Released: 1.9B Parameters Match 12B Models

Liquid AI today released the new-generation model family LFM3. SGLang has simultaneously launched Day-0 support, and the dev image is now available to pull. The entire family consists of small models at the 2B level, but several benchmark scores directly reach the 12B level.

LFM3 has a total of 6 instances, covering text, vision, audio, omni-modal, and decision-making. At its core is a new tri-hybrid architecture: 13 gated short convolution blocks, 5 sliding-window attention blocks, and 6 global attention blocks. The KV state grows with sequence length only in the 6 global attention layers, keeping long-context memory overhead very low.

Parameter efficiency is the most striking number this time. LFM3-1.7B scores 38.42 on WebArena-Verified and 41.29 on AutomationBench, both exceeding Qwen3.5-9B. LFM3-VL-2B is even more impressive: with 1.9B parameters, it averages 71.4 across 62 vision benchmarks, matching Gemma 4 12B, whose parameter count is 6 times larger.

The positioning of each of the six models:

• LFM3-1.7B: Text, 256K context. About 1 GB after 4-bit quantization, and it can run on a phone. Fine-tuning is possible on a single GPU.
• LFM3-VL-2B: Vision-language, adds video input with 2fps sampling. It can handle charts, math, object localization, and spatial relationships.
• LFM3-Audio-2B: Speech, 18 languages. The causal audio encoder supports processing while speaking, with first-token latency as low as 50ms.
• LFM3-Omni-2B: Omni-modal, vision + audio + language, 2.2B parameters.
• d1-2B / d1-Omni-2B: Decision models that output structured results and calibrated probabilities in a single forward pass.

Liquid AI designs models using hardware-in-the-loop architecture search, directly incorporating inference speed and memory usage into the constraints for architecture selection. LFM2 mainly optimized phone and laptop CPUs, while LFM3 expands to 14 device combinations: Galaxy S26 Ultra, OnePlus 15, Apple M4 Max, AMD Ryzen AI, Raspberry Pi 5, and various GPUs/CPUs/NPUs. Deployment constraints come before architecture, which is clearly different from the mainstream pretraining paradigm.

For long context, LFM3-1.7B uses 1.6 GB for the KV cache at 128K tokens, compared with 4.3 GB for Qwen3.5-4B. Liquid AI also ran 300,000 parallel agents to analyze the Enron email archive, demonstrating the feasibility of large-scale agent tasks.

The fine-tuning barrier is also low. In a wind turbine blade inspection case: about 3,000 public photos, SFT+GRPO fine-tuning of LFM3-VL-2B, completed in a few hours on a single A100. Defect classification accuracy rose from less than 10% to 67%, and detection F1 went from 0.003 to 0.11. The fine-tuned model can also answer image-related questions and generate inspection reports.

SGLang's Day-0 support is already available, but it has not yet entered an official release, so the dev image is required:

docker pull lmsysorg/sglang:dev-lfm3

Single-GPU launch:

sglang serve LiquidAI/LFM3-1.7B \
  --reasoning-parser auto \
  --tool-call-parser auto

The BF16 weights of both checkpoints are under 4 GB. LFM3 enables thinking mode by default, and SGLang splits <think> into reasoning_content. Tool calls are in XML format, and SGLang automatically parses them into standard tool_calls, so the OpenAI-compatible interface works directly.

The model card also clearly states the boundaries: small models are not suitable for agentic coding and complex multi-step reasoning. They are suitable for specialized scenarios such as on-device agents, voice assistants, and visual understanding.

Model cards: LFM3-1.7B, LFM3-VL-2B
SGLang Cookbook: https://docs.sglang.io/cookbook/autoregressive/LiquidAI/LFM3