Liquid AILFM3Lfm3VlForConditionalGeneration

LFM3 VL 2B — Hardware Requirements & GPU Compatibility

VisionChatFunctions

LFM3-VL-2B is Liquid AI's vision-language model for on-device visual understanding. It combines the 1.73-billion-parameter LFM3 language model with a new in-house 164-million-parameter vision encoder, about 1.9 billion parameters in all. It accepts text, images, and video, and is the first LFM vision model to take video as input. It can reason step by step before answering, which helps with math, charts, and multi-step visual questions, and it handles document and app-screen reading, object finding, and spatial reasoning. Context length is 128K tokens, input plus output. It is released under the LFM Open License v1.0, a custom license whose terms are in the repository, and was published in October 2026. With under 2 billion parameters it is small enough for a phone or any modern laptop, and it runs comfortably on a consumer GPU even at higher-precision weights.

1.9B params · dense128K contextOtherOct 20266.2K downloads180 likes51.0K quant downloads

Based on LFM3 1.7B Base

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Runs great on a 24 GB GPU

Runs from 8 GB (Q4_K_M · 125K)Best on 24 GB (Q4_K_M · ~423 tok/s · 125K)Macs from 8 GB (Q4_K_M · ~46 tok/s)

Other

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run LFM3 VL 2B

One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Ways to runOther ways

each opens step-by-step

Run in cloud

Runs on RTX 3060 12GB (9 GB headroom) · Q4_K_M

Generation speed: ~151 tok/s; generation speed
Cost per 1M output tokens: $0.11; per 1M output tokens

Deployon Vast.aiBest valuefrom $0.061/hr

Compare GPUs →

Run LFM3 VL 2B

2.3 GB VRAM at 16K context · fits the RTX 4090 (21.7 GB free)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/LiquidAI/LFM3-VL-2B-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and search liquid/lfm3-vl-2b, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get liquid/lfm3-vl-2b@q4_k_m
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf LiquidAI/LFM3-VL-2B-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/LiquidAI/LFM3-VL-2B-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf LiquidAI/LFM3-VL-2B-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve LiquidAI/LFM3-VL-2B --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~4 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model LiquidAI/LFM3-VL-2B-MLX-8bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model LiquidAI/LFM3-VL-2B-MLX-8bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$LiquidAI/LFM3-VL-2B-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
RTX 3060 12GBCheapestCheapest9 GB free	~151 tok/s	$0.11	Deploy on Vast.aifrom $0.061/hr (opens in new tab)	9 GB free · ~151 tok/s · $0.11 / 1M tokens
RTX 3060 Ti (8 GB)5 GB free	~188 tok/s	$0.11	Deploy on Vast.aifrom $0.072/hr (opens in new tab)	8 GB · 5 GB free · ~188 tok/s · $0.11 / 1M tokens
RTX 4060 (8 GB)5 GB free	~114 tok/s	$0.20	Deploy on Vast.aifrom $0.080/hr (opens in new tab)	8 GB · 5 GB free · ~114 tok/s · $0.20 / 1M tokens
RTX A4000 (16 GB)13 GB free	~188 tok/s	$0.14	Deploy on Vast.aifrom $0.092/hr (opens in new tab)	16 GB · 13 GB free · ~188 tok/s · $0.14 / 1M tokens
RTX 3070 (8 GB)5 GB free	~188 tok/s	$0.19	Deploy on RunPodfrom $0.13/hr (opens in new tab)	8 GB · 5 GB free · ~188 tok/s · $0.19 / 1M tokens
RTX 3090 (24 GB)21 GB free	~393 tok/s	$0.16	Deploy on RunPodfrom $0.22/hr (opens in new tab)	24 GB · 21 GB free · ~393 tok/s · $0.16 / 1M tokens
RTX A6000 (48 GB)45 GB free	~322 tok/s	$0.28	Deploy on RunPodfrom $0.33/hr (opens in new tab)	48 GB · 45 GB free · ~322 tok/s · $0.28 / 1M tokens
RTX 4090 (24 GB)21 GB free	~423 tok/s	$0.22	Deploy on RunPodfrom $0.34/hr (opens in new tab)	24 GB · 21 GB free · ~423 tok/s · $0.22 / 1M tokens

Compare GPUs →

How Much VRAM Does LFM3 VL 2B Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
1.1 GB	2.6 GB	—	0.7 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 125K context · ~624 tok/s
1.1 GB	2.7 GB	—	0.8 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 125K context · ~575 tok/s
1.2 GB	2.7 GB	—	0.8 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 125K context · ~551 tok/s
1.2 GB	2.7 GB	—	0.7 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 125K context · ~537 tok/s
1.2 GB	2.7 GB	—	0.9 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 125K context · ~537 tok/s
1.2 GB	2.8 GB	—	0.8 GB	3-bit small — 3-bit small quantization	Fits: 125K context · ~528 tok/s
1.3 GB	2.8 GB	—	0.9 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 125K context · ~520 tok/s
1.3 GB	2.9 GB	—	0.8 GB	3-bit medium — 3-bit medium quantization	Fits: 125K context · ~489 tok/s
1.4 GB	2.9 GB	—	1 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 125K context · ~482 tok/s
1.4 GB	2.9 GB	—	0.9 GB	3-bit large — 3-bit large quantization	Fits: 125K context · ~475 tok/s
1.4 GB	2.9 GB	—	1.1 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 125K context · ~458 tok/s
1.5 GB	3 GB	—	1.1 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Fits: 125K context · ~443 tok/s
1.5 GB	3 GB	—	1.1 GB	4-bit small — 4-bit small quantization	Fits: 125K context · ~443 tok/s
1.5 GB	3 GB	—	1.1 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Fits: 125K context · ~443 tok/s
1.6 GB	3.1 GB	—	1.1 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 125K context · ~423 tok/s
1.6 GB	3.1 GB	—	1.2 GB	4-bit large — 4-bit large quantization	Fits: 125K context · ~415 tok/s
1.7 GB	3.2 GB	—	1.3 GB	5-bit small — 5-bit small quantization	Fits: 125K context · ~381 tok/s
1.8 GB	3.3 GB	—	1.3 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Fits: 125K context · ~370 tok/s
2.0 GB	3.5 GB	—	1.4 GB	6-bit, very good — 6-bit quantization, very good quality	Fits: 125K context · ~331 tok/s
2.3 GB	3.8 GB	—	1.8 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Fits: 125K context · ~282 tok/s
4.2 GB	5.8 GB	—	3.5 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Fits: 125K context · ~155 tok/s

Selected Q4_K_M

4-bit medium quantization

2.3GBat 16K context

Weights: 1.2 GB
KV cache: 0.8 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · 21.8 GB free · ~423 tok/s

≥50 tok/s needs ≥119 GB/s · RTX 5090 ~752, RTX 3090 Ti ~423

Which Hardware Can Run LFM3 VL 2B?

Q4_K_M · 1.6 GB

LFM3 VL 2B (Q4_K_M) requires 1.6 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 3+ GB is recommended. Using the full 128K context window can add up to 6.2 GB, bringing total usage to 7.7 GB. 52 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 163 devices with unified memory can run LFM3 VL 2B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8 GB · RTX 4060Q4_K_M · 2.3 GB · ~114 tok/s · 125K2.3~114125K8 GB tier →

12 GB · RTX 3060 12GBQ4_K_M · 2.3 GB · ~151 tok/s · 125K2.3~151125K12 GB tier →

16 GB · RTX 5060 Ti 16GBQ4_K_M · 2.3 GB · ~188 tok/s · 125K2.3~188125K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 2.3 GB · ~423 tok/s · 125K2.3~423125K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 2.3 GB · ~54 tok/s · 125K2.3~54125KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 2.3 GB · ~247 tok/s · 125K2.3~247125KMacs →

Alternatives to LFM3 VL 2B

LFM3 family · 3 models →

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Moondream2vikhyatk	—	3 GB	~516	same size
Qwen 1 8BAlibaba	—	1.9 GB	~542	same size
JiRackUltra 1BCMSManhattan	—	1.8 GB	~461	same size
Gemma 3n E2B IT Litert LmGoogle	—	3.1 GB	~496	same size
Llm Jp 3.1 1.8B Instruct4llm-jp	—	2.2 GB	~360	same size
Helium 1 2Bkyutai	—	2 GB	~374	same size

Fine-tunes & variants(1)community builds of LFM3 VL 2B

D1 2B

Specifications

Publisher: Liquid AI
Family: LFM3
Parameters: 1.9B
Architecture: Lfm3VlForConditionalGeneration
Context Length: 128,000 tokens
Vocabulary Size: 128,000
Release Date: 2026-10-19
License: Other

Resources

Frequently Asked Questions

How much VRAM does LFM3 VL 2B need?

LFM3 VL 2B requires 1.6 GB of VRAM at Q4_K_M, or 4.2 GB at BF16. Full 128K context adds up to 6.2 GB (7.7 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 1.9B × 4.8 bits ÷ 8 = 1.2 GB

KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 6.5 GB (at full 128K context)

VRAM usage by quantization

Q4_K_M

1.6 GB

Q4_K_M + full context

7.7 GB

Learn more about VRAM estimation →

What's the best quantization for LFM3 VL 2B?

For LFM3 VL 2B, Q4_K_M (1.6 GB) offers the best balance of quality and VRAM usage. Q4_K_L (1.6 GB) provides better quality if you have the VRAM. The smallest option is IQ2_M at 1.1 GB.

VRAM requirement by quantization

IQ2_M

1.1 GB~62%

Q3_K_S

1.2 GB~77%

IQ4_XS

1.4 GB~87%

Q4_K_M ★

1.6 GB~89%

Q4_K_L

1.6 GB~90%

BF16

4.2 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run LFM3 VL 2B on a Mac?

Yes — MacBook Air 13" M3 (8 GB) and 145 other Macs can run LFM3 VL 2B. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 1.6 GB of usable unified memory (RAM minus macOS overhead).

Can I run LFM3 VL 2B locally?

Yes — LFM3 VL 2B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is LFM3 VL 2B?

At Q4_K_M, LFM3 VL 2B can reach ~3097 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~423 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = (bandwidth GB/s ÷ model GB) × efficiency

Example: NVIDIA B200 → 8000 ÷ 1.55 × 0.65 = ~3355 tok/s

Estimated speed at Q4_K_M (1.6 GB)

NVIDIA B200

~3355 tok/s

NVIDIA GeForce RTX 4090

~423 tok/s

NVIDIA B300

~3355 tok/s

AMD Instinct MI350X

~3097 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of LFM3 VL 2B?

At Q4_K_M, the download is about 1.15 GB. The full-precision BF16 version is 3.84 GB. The smallest option (IQ2_M) is 0.65 GB.

Which GPUs can run LFM3 VL 2B?

52 consumer GPUs can run LFM3 VL 2B at Q4_K_M. Top options include AMD Radeon RX 6700 XT, AMD Radeon RX 6800, AMD Radeon RX 6800 XT. 52 GPUs have plenty of headroom for comfortable inference.

Which devices can run LFM3 VL 2B?

163 devices with unified memory can run LFM3 VL 2B at Q4_K_M (1.6 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.