LFM2.5-24B-A2B VRAM requirements: can you run it locally?

LFM2.5-24B-A2B is Liquid AI's largest LFM2.5 model: a hybrid mixture-of-experts built for on-device agents. Its Hugging Face repo was created on October 12, 2026 (blog post October 13), it uses the LFM Open License v1.0, and Hugging Face lists 23,971,585,280 parameters.^[1][15] The model card states 2.45B active parameters per token, 40 layers (30 short-convolution blocks + 10 grouped-query attention), 64 experts with top-4 routing, and a 131,072-token context.^[2][3] Liquid reports, for example, 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B; these are developer-reported scores.^[2]

Because it is an MoE, only a small slice of the weights does work for each token, which is why it runs fast. But all 24B parameters still have to sit in memory, so memory, not speed, is what decides whether it runs for you.

Memory needed at each quantization

Quantization stores each weight in fewer bits. Smaller files fit on smaller GPUs, at some cost in quality, and the cost grows as you go below about 4 bits. File sizes below are the actual GGUF downloads on Hugging Face (1 GB = 1,000,000,000 bytes). Q4_K_M through BF16 are from Liquid's official GGUF repo; the 2-bit and 3-bit rows are Unsloth dynamic quants.^[4][5]

Ollama's default lfm2.5:24b tag is 15 GB, matching its 24b-a2b-q4_K_M tag. Its 24b-a2b-q4_0 tag is 14 GB, 24b-a2b-q8_0 is 25 GB, and 24b-a2b-bf16 is 48 GB.^[6]

Context length adds memory (the KV cache)

Only the 10 full-attention layers keep a key/value cache that grows with every token.^[3] The 30 convolution layers use a fixed short state that does not grow with context. With 8 KV heads and a head size of 64 (hidden size 2048 ÷ 32 heads), an FP16 cache takes 2 × 10 × 8 × 64 × 2 bytes = 20,480 bytes per token:

So the full 128K context needs only about 2.7 GB of FP16 cache on top of the weights — far cheaper than a dense 24B–33B model. An 8-bit or 4-bit KV cache shrinks these numbers further.

Which GPUs and Macs it fits on

Each cell compares the weights plus an 8K-token KV cache with the device's memory.^{[7][8][9][10][11][12][13]}

Column headers show weights + 8,192-token KV cache.

How to read it: "Fits" means weights + 8K cache use under 90% of memory. "Tight" means 90–100%: it may load, but the engine's own working memory or a longer context can push it over. "No" means it won't fit on the GPU alone. The 90% line is our own rule of thumb, not a vendor figure. On a Mac, macOS also needs some of the unified memory, so treat Mac rows as optimistic.

How to run it

The easiest way is Ollama. ollama run lfm2.5:24b pulls the 15 GB Q4_K_M build.^[6] For a specific quant, use tags such as lfm2.5:24b-a2b-q4_0 (14 GB) or lfm2.5:24b-a2b-q8_0 (25 GB).^[6] With llama.cpp or LM Studio, download a GGUF from Liquid's official list (for example LFM2.5-24B-A2B-Q4_K_M.gguf, 14.5 GB) or an Unsloth 2-/3-bit file for 12 GB cards.^[4][5]

If the model is bigger than your VRAM, these tools can keep some layers in system RAM, but every token then has to wait on the slower memory, so expect much lower speed.

What people report (third-party, not verified by us): one X account timed Liquid's QAD-Q4_0 (13.6 GB) at about 89 tok/s on an M2 Pro 32 GB, tying larger quants on their accuracy set.^[16] Another local-coding roundup ranked LFM2.5 24B third behind Qwen3.6 35B-A3B and Ornith 1.5 35B-A3B.^[17]

Cheapest GPU to rent for LFM2.5-24B-A2B

A 16 GB or 24 GB GPU runs the 4-bit build with room for context; 32 GB or 48 GB leaves headroom for Q8 or longer contexts. On-demand Secure Cloud prices listed on Runpod's pricing page:^[14]

Prices checked on 2026-10-27. They change often, so confirm on the provider's page. Rent on Runpod Referral link

Frequently asked questions

Can I run LFM2.5-24B-A2B on an RTX 3090 or RTX 4090?

Yes. Q4_K_M (14.5 GB) plus an 8K cache is about 14.7 GB, well under 24 GB.^{[4][3][9][10]} Q5_K_M (17.0 GB) and Q6_K (19.7 GB) also fit at short contexts; Q8_0 (25.5 GB) does not fit on 24 GB alone.

Can I run it on a 16 GB or 12 GB GPU?

On 16 GB (for example an RTX 5080), Q4_K_M fits with room for a moderate context.^[4][8] On 12 GB (for example an RTX 3060), use Unsloth's UD-Q3_K_XL (10.9 GB) or a 2-bit build; Q4_K_M does not fit.^[5][7]

How much VRAM does the full 128K context need?

At FP16, the KV cache for 131,072 tokens is about 2.7 GB, on top of the weights.^[3] With Q4_K_M that is about 17.2 GB before engine overhead, which still fits a single 24 GB card.

Why is the KV cache so small for a 24B model?

Only 10 of the 40 layers are full attention. The other 30 are short-convolution blocks with a fixed-size state, so they do not grow with context the way a dense transformer's KV cache does.^[3]

Can a Mac run LFM2.5-24B-A2B?

Yes. A Mac with 24 GB of unified memory can run Q4_K_M; 32 GB holds Q5_K_M or Q6_K comfortably; 64 GB can run Q8_0 (25.5 GB).^[4][13] Liquid also ships MLX 4-bit builds for Apple Silicon.^[2]

How we calculated

• Weights are the exact GGUF file sizes on Hugging Face.^[4][5] BF16 is the official BF16 GGUF.
• KV cache is calculated from config.json (10 full-attention layers, 8 KV heads, head size 64). We assume an FP16 cache.^[3]
• Device memory is from NVIDIA and Apple spec pages. We compare in decimal GB, so the results lean conservative.
• Not included: engine working buffers, the OS and display, and any vision or draft (DSpark) add-ons. Leave headroom.

Sources

All sources checked on 2026-10-27.

1. Hugging Face, LiquidAI/LFM2.5-24B-A2B (parameter count, license, repo date, downloads; checked October 27, 2026). https://huggingface.co/LiquidAI/LFM2.5-24B-A2B
2. Hugging Face, LFM2.5-24B-A2B model card (2.45B active, architecture, benchmarks, formats). https://huggingface.co/LiquidAI/LFM2.5-24B-A2B/blob/main/README.md
3. Hugging Face, config.json (layers, KV heads, head size, context, layer_types). https://huggingface.co/LiquidAI/LFM2.5-24B-A2B/blob/main/config.json
4. Hugging Face, LiquidAI/LFM2.5-24B-A2B-GGUF file list. https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-GGUF/tree/main
5. Hugging Face, unsloth/LFM2.5-24B-A2B-GGUF file list (2-bit and 3-bit sizes). https://huggingface.co/unsloth/LFM2.5-24B-A2B-GGUF/tree/main
6. Ollama, lfm2.5 tags. https://ollama.com/library/lfm2.5/tags
7. NVIDIA, GeForce RTX 3060 family (12 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3060-3060ti/
8. NVIDIA, GeForce RTX 5080 (16 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5080/
9. NVIDIA, GeForce RTX 3090 specs (24 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/
10. NVIDIA, GeForce RTX 4090 specs (24 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/
11. NVIDIA, GeForce RTX 5090 specs (32 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/
12. NVIDIA, RTX 6000 Ada Generation (48 GB). https://www.nvidia.com/en-us/products/workstations/rtx-6000/
13. Apple, MacBook Pro tech specs (unified memory options). https://www.apple.com/macbook-pro/specs/
14. Runpod, GPU pricing (checked October 23, 2026). https://www.runpod.io/pricing
15. Liquid AI blog, LFM2.5-24B-A2B (published October 13, 2026). https://www.liquid.ai/blog/lfm2-5-24b-a2b
16. @QuantDiffDotCom on X, QAD-Q4_0 timing on M2 Pro (third-party report, October 26, 2026). https://x.com/QuantDiffDotCom/status/2114707369091379341
17. @Manz on X, local coding model ranking including LFM2.5 24B (third-party, October 26, 2026). https://x.com/Manz/status/2114689457160913327