LFM3 Requirements
VRAM, RAM and Devices for All 6 Models (2026)
LFM3 from Liquid AI (October 2026): LFM3-1.7B needs about 1GB at 4-bit and 3.5GB in BF16. Memory for 6 models, 256K context, Japanese, license, setup.
LFM3-1.7B (1.73B parameters) needs about 1GB at 4-bit and roughly 3.5GB in BF16. Released by Liquid AI on October 27, 2026, all six LFM3 models are around 2B parameters, so their weights fit in roughly 1 to 4.4GB. An 8GB laptop, a Mac, or a recent smartphone is a realistic target. This article separates figures from the official model cards from estimates computed from parameter counts.
LFM3 requirements at a glance
| Model | Official params | BF16 (weights) | 8-bit (weights) | 4-bit (weights) | Minimum device (estimate) |
|---|---|---|---|---|---|
| LFM3-1.7B (text) | 1.73B | ~3.5GB | ~1.8GB | ~0.9-1GB (official: about 1GB) | 8GB RAM PC/Mac, phone, CPU-only |
| LFM3-VL-2B (vision) | 1.92B | ~3.8GB | ~1.9GB | ~1.0-1.2GB | 8GB RAM PC/Mac, recent phone |
| LFM3-Audio-2B | 2B (no detail officially) | ~4GB | ~2GB | ~1.0-1.3GB | 8GB RAM PC/Mac |
| LFM3-Omni-2B | 2.17B | ~4.3GB | ~2.2GB | ~1.2-1.5GB | 16GB RAM recommended for video |
| d1-2B | 1.92B | ~3.8GB | ~1.9GB | ~1.0-1.2GB | 8GB RAM PC/Mac |
| d1-omni-2B | Not stated (2B) | ~4.3GB | ~2.2GB | ~1.2-1.5GB | 16GB RAM recommended |
Memory values are weight-only estimates (parameters x bits / 8). The only official figure is about 1GB for LFM3-1.7B at 4-bit. Budget 1.2 to 1.5 times the weights at runtime for KV cache and buffers, more for long context or video.
To check whether your GPU or Mac can run a model, the VRAM calculator (free, no sign-up) estimates required VRAM from model, quantization, and context length. LFM3-1.7B is selectable in the model list. The tool covers text models only, not audio, omni, or d1 media processing.
What was released
LFM3 is a new family of open-weight models announced on October 27, 2026. The text model has a Base variant pretrained on 22T tokens. The architecture is a tri-hybrid: 13 gated short convolution layers, 5 sliding-window attention layers, and 6 global attention layers across 24 layers, with kernel size raised from 3 to 7. The official KV cache size is not stated.
Model details
LFM3-1.7B has 1.73B parameters, a 256K context, and a 128,000-token vocabulary. Vendor-reported scores include BFCL V4 71.39, ToolSandbox 80.36, IFEval 96.66, GPQA Diamond 63.37, LiveCodeBench v6 73.46, and WebArena-Verified 38.42 (versus 3.02 for LFM2.5-2.6B). These are Liquid AI's own numbers.
LFM3-VL-2B totals 1.92B (1.73B language model, 164M vision encoder, 25M projector) with a 128K context and video input. Vendor-reported average over 62 vision benchmarks is 71.4. LFM3-Audio-2B has a 32K context, 16kHz input and 24kHz output; reported ASR WER is 6.23 in English. LFM3-Omni-2B has 2.17B parameters, a 32K context, and takes text, image, video, and audio. d1-2B (1.92B) and d1-omni-2B are decision models that return calibrated probabilities for many questions in one forward pass; d1-omni-2B specs are not stated officially.
Changes from LFM2.5
| Item | LFM2.5-2.6B | LFM3-1.7B |
|---|---|---|
| Parameters | ~2.69B | 1.73B (about 35% fewer) |
| Layers | 30 (22 conv + 8 GQA) | 24 (13 conv + 5 SWA + 6 global) |
| Context | 128K | 256K |
| Languages | 16 (incl. Japanese) | 16 (incl. Japanese) |
| 4-bit memory | Under 2.5GB (official) | About 1GB (official) |
| License | LFM1.0 | LFM Open License v1.0 |
See our LFM2.5-2.6B requirements guide and the Liquid AI Japanese models guide.
Running on your hardware
Apple Silicon: a 4-bit MLX build exists for LFM3-1.7B, VL, and Audio. NVIDIA: BF16 works with transformers, vLLM, and SGLang; FP8 and NVFP4 builds exist for LFM3-1.7B. Weights fit in 4GB, but 6 to 8GB is more comfortable. CPU: GGUF with llama.cpp. Phones: Liquid AI benchmarks against iPhone 17 Pro and Galaxy S26 Ultra, with no specific RAM requirement stated.
Quick start
pip install -U transformers accelerate
from transformers import pipeline
generator = pipeline("text-generation", model="LiquidAI/LFM3-1.7B", trust_remote_code=True)
vllm serve "LiquidAI/LFM3-1.7B"
# Command forms below are standard but unverified by us
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF
ollama run hf.co/LiquidAI/LFM3-1.7B-GGUFRecommended sampling for general use is temperature 0.1, top_k 50, repetition_penalty 1.1. VL needs transformers 5.21.0 or later, Audio and Omni 5.20.2 or later.
License
All models use LFM Open License v1.0. Commercial use is licensed only for users or entities with annual revenue under US$10 million. Larger organizations are not licensed for commercial use under this agreement. Read the original text before adopting it.
Comparison with similar small models
The Omni model card compares against Gemma 4 E2B/E4B and Phi-4 multimodal, but we did not verify third-party numbers, so we make no ranking claims. See our Gemma 4 hardware requirements guide. The main practical difference is the revenue-based license cap compared with Apache 2.0 models.
Fixing memory errors
- Shorten the context length to shrink the KV cache
- Use a lower-bit quantization such as Q4_K_M or MLX 4-bit
- Reduce video frames or fps, and lower image resolution
- Check that your transformers version meets the minimum
FAQ
How much memory does LFM3-1.7B need?
Officially about 1GB at 4-bit. BF16 weights are about 3.5GB (estimate), plus runtime overhead.
Does LFM3 support Japanese?
Yes. Japanese is among the 16 text languages and the speech locales of the Audio and Omni models. No Japanese-specific official benchmark was found.
Can I use LFM3 commercially?
Under LFM Open License v1.0, only if annual revenue is under US$10 million. Check the license text.
Can it run on a phone?
It is designed for edge devices and Liquid AI compares on iPhone 17 Pro and Galaxy S26 Ultra, but no specific device requirement is stated.
What are d1 models?
Decision models that output calibrated probabilities for several questions in one forward pass instead of free text.
Related free tools (no sign-up, instant results)
Feel free to contact us
Contact Us