LFM2.5-24B-A2B hardware requirements: VRAM, GPU and RAM
- Page updated
- Sources checked
- GGUF sizes and headers checked October 22, 2026
- Capacity estimates, labeled
- Jump to the checker ↓
LFM2.5-24B-A2B is Liquid AI's 24-billion-parameter open-weight reasoning model, released October 13, 2026. It is a text-only mixture of experts with 2.45B parameters active per token, built for agents and tool calling on your own device, and only 10 of its 40 layers keep a KV cache.
Short answer: with llama.cpp the official 4-bit Q4_K_M (14.5 GB) needs about 17 GB in total at a 32K context (RigCheck estimate). It fits a 24 GB card (RTX 3090, 4090) or a 32 GB Mac. A 16 GB card holds Unsloth's 3-bit UD-Q3_K_XL fully at 32K, or Q4_K_M with 7 of its 40 layers in system RAM (slower). Without a GPU, 32 GB of RAM holds Q6_K. Ollama's lfm2.5:24b is a 15 GB download.
Can my hardware run LFM2.5-24B-A2B?
Open in a full page ↗Loading private checker…
RigCheck allocation estimate. Input embeddings stay in system RAM even when all decoder layers fit on the GPU. Their per-format sizes come from the GGUF headers of the linked files (read October 22, 2026); the model ties its output to the input embeddings, so there is no separate output tensor, and when the tied tensor crosses CPU and GPU buffer types the estimate also reserves an output copy (unified-memory estimates keep this conservative allowance, though a backend may reuse mapped storage; CPU-only runs do not add it). We add a RigCheck allowance of 0.5 GB for host runtime/staging buffers, separate from the 8 GB OS reserve. Decoder weights are placed per layer from each file's own block sizes in its GGUF header (the two leading dense layers are much lighter than the 38 mixture-of-experts layers); the output layer sits on the last GPU, compute buffers on the first. Every file has 40 GGUF blocks and no MTP block, so full offload is -ngl 41. Commands disable automatic fitting with --fit off and pin decoder layers and the output layer with -ngl, using --tensor-split for multiple cards. This is a mixture-of-experts model: keeping only the expert weights in system RAM (--n-cpu-moe) is another split, whose memory and speed this checker does not estimate. Actual tensor placement and buffer peaks can still differ; confirm runtime allocation before filling memory.
- Developer
- Liquid AI Official
- Params
- 24.0B total, 2.45B active per token Official
- Context
- 131,072 tokens Official
- Input
- text only Official
- License
- LFM Open License v1.0 Official
- KV cache
- 20 KiB per token at f16 RigCheck calculation
- llama.cpp
- Liquid AI publishes GGUF files for llama.cpp Official
- Ollama
- Ollama has lfm2.5:24b Official
How much VRAM does LFM2.5-24B-A2B need?
The GGUF files as Hugging Face lists them (decimal GB, read October 22, 2026), and the total memory each needs at a 32K context with an f16 KV cache (RigCheck estimate). The official weights are BF16, 48.0 GB in safetensors.
| File | File size | Needs at 32K | Source |
|---|---|---|---|
| BF16 | 48.0 GB | 51.0 GB | Liquid AI |
| Q8_0 | 25.5 GB | 28.3 GB | Liquid AI |
| Q6_K | 19.7 GB | 22.4 GB | Liquid AI |
| Q5_K_M | 17.0 GB | 19.7 GB | Liquid AI |
| Q4_K_M | 14.5 GB | 17.2 GB | Liquid AI |
| QAD-Q4_0 | 13.6 GB | 16.3 GB | Liquid AI |
| UD-Q3_K_XL | 10.9 GB | 13.7 GB | Unsloth |
| UD-IQ3_XXS | 9.0 GB | 11.7 GB | Unsloth |
| UD-Q2_K_XL | 8.0 GB | 10.7 GB | Unsloth |
| UD-IQ2_XXS | 7.5 GB | 10.1 GB | Unsloth |
QAD-Q4_0 is Liquid AI's quantization-aware distilled 4-bit file, trained to recover the accuracy lost to Q4_0 (Liquid AI's GGUF card). Liquid AI's repository also has an F16 file (48.0 GB), and Unsloth's has its own Q8_0 and BF16, UD-Q4_K_XL to UD-Q6_K_XL and UD-IQ2_M (7.5 GB); the checker leaves those out. Liquid AI's model card also lists an MLX 4-bit build for Apple Silicon, whose memory this checker does not estimate. Ollama's lfm2.5:24b package is 15 GB (q4_K_M); it also lists q4_0 (14 GB), q8_0 (25 GB) and bf16 (48 GB).
What hardware can run LFM2.5-24B-A2B?
RigCheck estimate at a 32K context with an f16 KV cache, from the file sizes above and the assumptions under the checker. "Largest" means the largest of the checker's ten files.
| Hardware | Largest file that fits fully | With layers in system RAM (slower) |
|---|---|---|
| No GPU + 32 GB RAM | Q6_K (all in RAM, CPU only) | no larger file |
| RTX 4060 (8 GB) + 32 GB RAM | none | Q8_0 (32 of 40 layers, 21 GB in RAM) |
| RTX 3060 (12 GB) + 32 GB RAM | UD-IQ3_XXS | Q8_0 (27 of 40 layers, 18 GB in RAM) |
| RTX 5060 Ti (16 GB) + 32 GB RAM | UD-Q3_K_XL | Q8_0 (21 of 40 layers, 14 GB in RAM) |
| RX 9070 XT (16 GB) + 32 GB RAM | UD-Q3_K_XL | Q8_0 (21 of 40 layers, 14 GB in RAM) |
| RTX 4090 (24 GB) + 32 GB RAM | Q6_K | Q8_0 (10 of 40 layers, 6 GB in RAM) |
| RX 7900 XTX (24 GB) + 32 GB RAM | Q6_K | Q8_0 (10 of 40 layers, 6 GB in RAM) |
| RTX 5090 (32 GB) + 32 GB RAM | Q8_0 | BF16 (19 of 40 layers, 23 GB in RAM) |
| Mac, 16 GB unified | none | not applicable (unified memory) |
| Mac, 24 GB unified | UD-Q3_K_XL | not applicable (unified memory) |
| Mac, 32 GB unified | Q6_K | not applicable (unified memory) |
| Mac, 64 GB unified | BF16 | not applicable (unified memory) |
Capacity only, not tested setups. The KV cache is small for this model (0.7 GB at 32K, 2.7 GB at the 128K maximum), so context length changes the answer less than on most models of this size.
Measured LFM2.5-24B-A2B on your own hardware, or tried and it failed? Send us the numbers: paste the benchmark output or log into the form (or email it). After checking it we add it to this page, labeled as a community report and shared under CC BY 4.0.
LFM2.5-24B-A2B specs
| Parameters | 24.0B total, 2.45B active per token; 64 experts per MoE block, 4 used per token Model card |
|---|---|
| Architecture | 40 layers: 30 double-gated short-convolution blocks and 10 grouped-query attention blocks; text only Model card |
| Context length | 131,072 tokens Model card |
| KV cache | 20 KiB per token at f16 (10 attention layers × 8 KV heads × 64 dims): 0.7 GB at 32K, 2.7 GB at 128K; plus a fixed 0.5 MB convolution state RigCheck calculation |
| Input | text only; always reasons before it answers Model card |
| License | LFM Open License v1.0: free to use, modify and redistribute; commercial use is licensed only while you (with your affiliates) have under US$10 million in annual revenue License |
| Runtimes | Liquid AI names Transformers, vLLM, SGLang, llama.cpp, MLX and LM Studio; Liquid AI publishes GGUF files for llama.cpp and Ollama has lfm2.5:24b (checked October 22, 2026) |
FAQ
How much VRAM does LFM2.5-24B-A2B need?
By our estimate about 17 GB at a 32K context for the official 4-bit Q4_K_M (14.5 GB file), so a 24 GB card holds it, and about 14 GB for Unsloth's 3-bit UD-Q3_K_XL (10.9 GB), which fits a 16 GB card. Liquid AI does not publish a memory table.
Can LFM2.5-24B-A2B run on a 16 GB GPU like the RTX 5060 Ti, 5080 or RX 9070 XT?
Yes, by our estimate: UD-Q3_K_XL fits fully at a 32K context, and the official Q4_K_M runs with 7 of its 40 layers in system RAM, which is slower. With an 8K context Q4_K_M still keeps 5 layers in system RAM; most of its 17 GB is the file itself, not the KV cache.
Can LFM2.5-24B-A2B run on an 8 GB or 12 GB GPU?
On a 12 GB card the 3-bit UD-IQ3_XXS (9.0 GB) fits fully at 32K; larger files run with some layers in system RAM. On an 8 GB card none of the files fits fully; with 32 GB of RAM, part of the layers run from system RAM (our estimate), which is slower.
Can I run LFM2.5-24B-A2B on a Mac?
Yes. By our conservative estimate a Mac with 24 GB holds UD-Q3_K_XL and one with 32 GB holds Q6_K at a 32K context; a 16 GB Mac holds none of these files. Use llama.cpp (Metal), Ollama, or Liquid AI's MLX 4-bit build. Liquid AI measured 184 tokens per second of decoding on an Apple M5 Max (memory size not stated) with llama.cpp, Q4_K_M and a 4K-token input; it is that machine's figure, not an estimate for other Macs.
Can I run LFM2.5-24B-A2B without a GPU?
Yes, with llama.cpp's CPU build: by our estimate 32 GB of RAM holds Q6_K at a 32K context and 24 GB holds UD-Q3_K_XL; with 16 GB none of the files fits after the 8 GB we keep for the OS. Liquid AI's model card describes the GGUF files as optimized for CPU inference; we found no published CPU speed.
Is LFM2.5-24B-A2B on Ollama?
Yes: ollama run lfm2.5:24b downloads a 15 GB q4_K_M package; Ollama lists the model's context window as 128K. Ollama also lists lfm2.5:24b-a2b-q4_0 (14 GB), q8_0 (25 GB) and bf16 (48 GB) (checked October 22, 2026). Plain lfm2.5 is the smaller 8B-A1B model.
Can I use LFM2.5-24B-A2B commercially?
Under the LFM Open License v1.0, commercial use is licensed only while you, counted together with your affiliates, have under US$10 million in annual revenue; above that, commercial use is not covered by this licence (contact Liquid AI). Qualified non-profits (tax-exempt organisations run for charitable, religious, scientific, educational, literary or public-safety-testing purposes) above the threshold are covered only for non-commercial or research use. This is a summary, not legal advice; read the licence.
Sources
- Hugging Face: LiquidAI/LFM2.5-24B-A2B (model card, config.json, LFM Open License v1.0, file sizes) and Liquid AI's release post (October 13, 2026)
- Hugging Face: LiquidAI/LFM2.5-24B-A2B-GGUF (official GGUF sizes and headers, llama.cpp commands, QAD-Q4_0)
- Hugging Face: unsloth/LFM2.5-24B-A2B-GGUF (smaller GGUF sizes and headers)
- Ollama library: lfm2.5 tags and Ollama's GPU support list (checked October 22, 2026)
- llama.cpp: src/llama-arch.cpp (the lfm2moe architecture), src/llama-hparams.cpp (convolution state size), src/llama-memory-recurrent.cpp (state follows each layer device) and tools/server/README.md
- Usable-memory reserves, buffers and the offload model are RigCheck assumptions, stated under the checker.