Liquid AILFM2.5Lfm2MoeForCausalLM

LFM2.5 24B A2B — Hardware Requirements & GPU Compatibility

ChatFunctions

LFM2.5-24B-A2B is Liquid AI's largest and final model in the LFM2.5 family, a text-only reasoning model with 24 billion total parameters and about 2.45 billion active per token. It is aimed at fast, private agents that run on-device, and its hybrid design mixes short-convolution blocks with a smaller number of attention layers. Because only a couple of billion parameters fire per token, decoding is quick; the card cites 184 tokens per second on an Apple M5 Max. A Q4 GGUF takes roughly 14 GB, so it fits a 16 GB GPU or a 24 GB Mac comfortably. The context length is 131,072 tokens. The model always thinks before answering, and the card highlights agentic and tool-use performance, comparing it favorably with the larger Qwen3.6-35B-A3B. It supports 16 languages, including English, Chinese, Japanese, Spanish, and Arabic. It ships under the LFM 1.0 license, a custom license rather than Apache or MIT, so check its terms before commercial use.

24.0B params · MoE2.5B active131K contextOtherOct 20269.4K downloads431 likes122.6K quant downloads

Based on LFM2.5 24B A2B 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 (IQ2_XS · 16K)Best on 24 GB (Q4_K_M · ~221 tok/s · 128K)Macs from 32 GB (~38 tok/s)

Other

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run LFM2.5 24B A2B

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Other ways

each opens step-by-step

Run in cloud

Fits on RTX A5000 (24 GB) (9 GB headroom) · Q4_K_M

Generation speed: ~191 tok/s; generation speed
Cost per 1M output tokens: $0.23; per 1M output tokens

Deployon RunPodBest optionfrom $0.16/hr

Compare GPUs →

Run LFM2.5 24B A2B

15.0 GB VRAM at 16K context · fits the RTX 4090 (9.0 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 lfm2.5:24b
Official Ollama library build. For the exact Q4_K_M file from LiquidAI/LFM2.5-24B-A2B-GGUF: ollama run hf.co/LiquidAI/LFM2.5-24B-A2B-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 paste https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-GGUF@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/LFM2.5-24B-A2B-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/LFM2.5-24B-A2B-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/LFM2.5-24B-A2B-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/LFM2.5-24B-A2B --max-model-len 16384
Runs the original ~53 GB weights, not the Q4_K_M GGUF.

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/LFM2.5-24B-A2B-MLX-4bit
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/LFM2.5-24B-A2B-MLX-4bit

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/LFM2.5-24B-A2B-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 A5000 (24 GB)CheapestCheapest9 GB free	~191 tok/s	$0.23	Deploy on RunPodfrom $0.16/hr (opens in new tab)	24 GB · 9 GB free · ~191 tok/s · $0.23 / 1M tokens
RTX 3090 (24 GB)9 GB free	~213 tok/s	$0.24	Deploy on Vast.aifrom $0.18/hr (opens in new tab)	24 GB · 9 GB free · ~213 tok/s · $0.24 / 1M tokens
RTX 4000 Ada (20 GB)5 GB free	~115 tok/s	$0.47	Deploy on Vast.aifrom $0.20/hr (opens in new tab)	20 GB · 5 GB free · ~115 tok/s · $0.47 / 1M tokens
RTX 4000 Ada (20 GB)5 GB free	~115 tok/s	$0.48	Deploy on RunPodfrom $0.20/hr (opens in new tab)	20 GB · 5 GB free · ~115 tok/s · $0.48 / 1M tokens
RTX 3090 (24 GB)9 GB free	~213 tok/s	$0.29	Deploy on RunPodfrom $0.22/hr (opens in new tab)	24 GB · 9 GB free · ~213 tok/s · $0.29 / 1M tokens
RTX A5000 (24 GB)9 GB free	~191 tok/s	$0.34	Deploy on Vast.aifrom $0.23/hr (opens in new tab)	24 GB · 9 GB free · ~191 tok/s · $0.34 / 1M tokens
RTX PRO 4000 (24 GB)9 GB free	~176 tok/s	$0.43	Deploy on Vast.aifrom $0.27/hr (opens in new tab)	24 GB · 9 GB free · ~176 tok/s · $0.43 / 1M tokens
RTX A6000 (48 GB)33 GB free	~191 tok/s	$0.48	Deploy on RunPodfrom $0.33/hr (opens in new tab)	48 GB · 33 GB free · ~191 tok/s · $0.48 / 1M tokens

Compare GPUs →

How Much VRAM Does LFM2.5 24B A2B Need?

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

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
7.1 GB	7.6 GB	9.6 GB	7.5 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 128K context · ~303 tok/s
7.7 GB	8.2 GB	10.2 GB	7.1 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 128K context · ~294 tok/s
8.0 GB	8.5 GB	10.5 GB	7.2 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 128K context · ~290 tok/s
8.6 GB	9.1 GB	11.1 GB	7.5 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 128K context · ~283 tok/s
9.8 GB	10.3 GB	12.3 GB	9 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 128K context · ~268 tok/s
10.1 GB	10.6 GB	12.6 GB	8.2 GB	2-bit · small K-quant — 2-bit small K-quant	Fits: 128K context · ~265 tok/s
10.4 GB	10.9 GB	12.9 GB	9.9 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 128K context · ~262 tok/s
10.7 GB	11.2 GB	13.2 GB	10.4 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 128K context · ~259 tok/s
10.7 GB	11.2 GB	13.2 GB	8.8 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 128K context · ~259 tok/s
11.0 GB	11.5 GB	13.5 GB	10.4 GB	3-bit small — 3-bit small quantization	Fits: 128K context · ~256 tok/s
11.3 GB	11.8 GB	13.8 GB	10.5 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 128K context · ~253 tok/s
12.2 GB	12.7 GB	14.7 GB	11.5 GB	3-bit medium — 3-bit medium quantization	Fits: 128K context · ~244 tok/s
12.4 GB	13 GB	15 GB	13.6 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 128K context · ~241 tok/s
12.8 GB	13.3 GB	15.3 GB	12.4 GB	3-bit large — 3-bit large quantization	Fits: 128K context · ~239 tok/s
13.3 GB	13.9 GB	15.9 GB	12.8 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 128K context · ~233 tok/s
13.9 GB	14.5 GB	16.5 GB	13.6 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Fits: 128K context · ~228 tok/s
13.9 GB	14.5 GB	16.5 GB	15.1 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Fits: 128K context · ~228 tok/s
13.9 GB	14.5 GB	16.5 GB	13.7 GB	4-bit small — 4-bit small quantization	Fits: 128K context · ~228 tok/s
14.8 GB	15.4 GB	17.4 GB	14.5 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 128K context · ~221 tok/s
15.2 GB	15.7 GB	17.7 GB	16.4 GB	4-bit large — 4-bit large quantization	Fits: 128K context · ~219 tok/s
15.4 GB	16 GB	18 GB	16.5 GB	5-bit · legacy — 5-bit legacy quantization	Fits: 128K context · ~217 tok/s
16.9 GB	17.5 GB	19.5 GB	18 GB	5-bit · legacy+ — 5-bit legacy quantization with offset	Fits: 128K context · ~206 tok/s
16.9 GB	17.5 GB	19.5 GB	16.5 GB	5-bit small — 5-bit small quantization	Fits: 128K context · ~206 tok/s
17.6 GB	18.1 GB	20.1 GB	17 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Fits: 128K context · ~202 tok/s
20.2 GB	20.7 GB	22.8 GB	19.7 GB	6-bit, very good — 6-bit quantization, very good quality	Fits: 128K context · ~186 tok/s
24.4 GB	24.9 GB	27 GB	25.5 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 24.6 GB
48.4 GB	48.9 GB	50.9 GB	48 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 48.6 GB

Selected Q4_K_M

4-bit medium quantization

15.0GBat 16K context

Weights: 14.4 GB
KV cache: 0.3 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · 9.0 GB free · ~221 tok/s

≥50 tok/s needs ≥131 GB/s · RTX 5090 ~285, RTX 3090 Ti ~221

Which Hardware Can Run LFM2.5 24B A2B?

Q4_K_M · 14.8 GB

LFM2.5 24B A2B (Q4_K_M) requires 14.8 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 20+ GB is recommended. Using the full 131K context window can add up to 2.5 GB, bringing total usage to 17.4 GB. 26 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 5080. 125 devices with unified memory can run LFM2.5 24B A2B, including NVIDIA DGX H100, NVIDIA DGX A100 640GB, Mac Mini M5 Pro (24 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8 GB · RTX 4060IQ2_XS · 7.8 GB · ~151 tok/s · 16K7.8~15116K8 GB tier →

12 GB · RTX 3060 12GBIQ3_M · 11.4 GB · ~140 tok/s · 32K11.4~14032K12 GB tier →

16 GB · RTX 5060 Ti 16GBQ4_K_M · 15 GB · ~135 tok/s · 32K15~13532K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 15 GB · ~221 tok/s · 128K15~221128K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 15 GB · ~38 tok/s · 128K15~38128KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 15 GB · ~84 tok/s · 128K15~84128KMacs →

Alternatives to LFM2.5 24B A2B

LFM2.5 family · 14 models →

Similar models

4tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Gemma 4 26B A4B ITGoogle	105.4	16.5 GB	~171	same size
Magistral Small 2509Mistral AI	85.5	18.1 GB	~43	same size
Ariarhymes-ai	—	20.2 GB	~168	same size
EuroLLM 22B Instruct 2512utter-project	—	17.5 GB	~46	same size

Fine-tunes & variants(1)community builds of LFM2.5 24B A2B

LFM2.5 24B A2B ONNX

Specifications

Publisher: Liquid AI
Family: LFM2.5
Parameters: 24.0B
Active parameters: 2.5B per token
Architecture: Lfm2MoeForCausalLM
Context Length: 131,072 tokens
Vocabulary Size: 128,000
Release Date: 2026-10-12
License: Other

Resources

Frequently Asked Questions

How much VRAM does LFM2.5 24B A2B need?

LFM2.5 24B A2B requires 14.8 GB of VRAM at Q4_K_M, or 48.4 GB at BF16. Full 131K context adds up to 2.5 GB (17.4 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 24.0B × 4.8 bits ÷ 8 = 14.4 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 ≈ 3 GB (at full 131K context)

VRAM usage by quantization

Q4_K_M

14.8 GB

Q4_K_M + full context

17.4 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run LFM2.5 24B A2B?

Yes, at Q6_K (20.2 GB) or lower. Higher quantizations like Q8_0 (24.4 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for LFM2.5 24B A2B?

For LFM2.5 24B A2B, Q4_K_M (14.8 GB) offers the best balance of quality and VRAM usage. Q4_K_L (15.2 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 7.1 GB.

VRAM requirement by quantization

IQ2_XXS

7.1 GB~53%

IQ3_S

10.7 GB~75%

Q3_K_L

12.8 GB~86%

Q4_K_M ★

14.8 GB~89%

Q5_0

15.4 GB~90%

BF16

48.4 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run LFM2.5 24B A2B on a Mac?

Yes, but only at lower quantizations. The smallest Mac that can run LFM2.5 24B A2B is iPad Pro M5 13" (16 GB) at IQ2_XXS; 136 of the 156 Macs we list can run it at some quantization. For Q4_K_M (14.8 GB) you need a Mac with more unified memory.

Can I run LFM2.5 24B A2B locally?

Yes — LFM2.5 24B A2B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is LFM2.5 24B A2B?

At Q4_K_M, LFM2.5 24B A2B can reach ~120 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~221 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)

Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.

Example: NVIDIA B200 → 1.5 GB active ÷ (8000 × 0.65) = 0.292 ms, plus 40 layers × 0.055 ms = 2.200 ms, so 1000 ÷ 2.492 ms = ~401 tok/s

Estimated speed at Q4_K_M (14.8 GB)

NVIDIA B200

~401 tok/s

NVIDIA GeForce RTX 4090

~221 tok/s

NVIDIA B300

~401 tok/s

NVIDIA GH200 Grace Hopper Superchip

~374 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 LFM2.5 24B A2B?

At Q4_K_M, the download is about 14.38 GB. The full-precision BF16 version is 47.94 GB. The smallest option (IQ2_XXS) is 6.59 GB.

Which GPUs can run LFM2.5 24B A2B?

26 consumer GPUs can run LFM2.5 24B A2B at Q4_K_M. Top options include AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090, NVIDIA GeForce RTX 3090 Ti, AMD Radeon RX 6800. 7 GPUs have plenty of headroom for comfortable inference.

Which devices can run LFM2.5 24B A2B?

125 devices with unified memory can run LFM2.5 24B A2B at Q4_K_M (14.8 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Asus ROG Flow Z13 (2025, Ryzen AI Max+ 395, 128 GB), Beelink GTR9 Pro (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.