How to Run LFM2.5-24B-A2B Locally: Only 2B Active, Still Needs 14.5 GB of RAM

LFM2.5-24B-A2B is Liquid AI's new 24B mixture-of-experts model, and it touches only 2.45B of its parameters per token. Liquid says it decodes at 184 tokens/s on an Apple M5 Max. You can run the GGUF locally in Ollama or llama.cpp with one command. It came out on October 13 and the GGUF repo already shows 17,807 downloads.

What the speed numbers leave out: all 24B weights still have to sit in memory. The Q4_K_M file is 14.5 GB. The "2B" helps your tokens per second and does nothing for your RAM.

Below: what changed from LFM2-24B-A2B, where the benchmarks hold up and where they don't, memory per quant, the run commands, and the license clause that matters if your company is big. I haven't run it myself yet. Every number comes from a linked source.

What LFM2.5-24B-A2B actually is

Liquid released it on October 13, 2026 (blog). It's the largest model in the LFM2.5 family, and the card calls it the final one. Specs from the model card:

• 24.0B total parameters, 2.45B active
• 40 layers: 30 double-gated short convolution blocks plus 10 GQA attention layers
• 64 experts per MoE block, top-4 routing
• 131,072 token context
• 128K vocabulary
• ~33T pretraining tokens (the same card also lists a 36T "training budget"), then 2T tokens of mid-training at 32K context and 300B tokens at 131K
• Text only. It is a reasoning model and always opens with a <think> block

People are already tripping over the text-only part. Someone opened a discussion titled "Great model, but useless for me without mmproj". There is no mmproj to ship, because there is no vision tower. If you want images, the LFM2.5 VL models stop at LFM2.5-VL-3B (model library).

What changed vs LFM2-24B-A2B

If you ran the earlier LFM2-24B-A2B (it's in Ollama as lfm2:24b), the architecture is the same: 40 layers, 64 experts, top-4, same conv/GQA hybrid. Per Liquid's blog, the differences are:

The rest of the LFM2.5 text lineup is small: 230M, 350M, 1.2B (Instruct, Thinking, JP), 2.6B, and the LFM2.5-8B-A1B MoE with 128K context. This 24B is the only one that tries to play in the 30B-class MoE league.

Where the benchmarks hold up

Liquid compares it against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning (30B-A3B). A selection from the card:

My read: this is a tool-calling and instruction-following model. It wins all three instruction-following benchmarks (IFEval 96.75, IFBench, Multi-IF) and five of the seven agentic ones, including LiveClaw at 48.09 vs Qwen's 38.19. It loses PinchBench and WebArena Verified to Qwen. Calling BFCLv4 a win is generous too: 65.01 vs 65.00 is a tie.

The losses are not close. Qwen3.6-35B-A3B is 15 points ahead on SWE-Bench Verified and 9 ahead on GPQA Diamond. The AA-Omniscience score of -24.90 (it penalizes confident wrong answers) is roughly level with Qwen, and miles behind Nemotron's -3.58. So if you were hoping for "mid-size knowledge in a 2B-speed package", the lab's own table says knowledge is the weak spot. Use it to follow a schema and call your functions.

Keep in mind these are lab-published numbers with per-benchmark settings (mostly temperature 0.6, vLLM, specific turn limits). I haven't found an independent reproduction yet.

How much memory LFM2.5-24B-A2B needs per quant

This is the question people are already asking in the repo (one discussion is titled "LFM2.5-24B A2B, VRAM Requirements"). File sizes from the GGUF repo. Budget for the file plus the KV cache, the OS, and whatever else you have open.

The "where it fits" column is my rule of thumb from the file sizes, not something Liquid publishes. Liquid's blog only says it uses "the least memory" among the MoEs it compared on edge hardware, without a GB figure.

The KV cache helps a bit. Only 10 of the 40 layers are attention layers, so it should grow much more slowly with context than in a pure transformer of the same depth, and long contexts cost less than you'd expect. Still, don't set 131K context on a 16 GB laptop and act surprised.

About the Q4_0 file: it's a Quantization-Aware Distillation checkpoint, which the repo says was trained "to recover the accuracy lost to Q4_0 quantization". Nice idea, but there's already a thread asking for the actual accuracy numbers, and so far Liquid hasn't posted any. Until they do, I'd treat Q4_K_M as the default and the QAD file as the one to A/B test.

Run LFM2.5-24B-A2B with llama.cpp, Ollama or LM Studio

The llama-cli commands are from the GGUF README. The server, Ollama and LM Studio lines are the standard Hugging Face "Use this model" snippets for the repo.

llama.cpp, OpenAI-compatible server:

llama-server -hf LiquidAI/LFM2.5-24B-A2B-GGUF:Q4_K_M

llama.cpp in the terminal, as the README gives it:

llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF \
    --temp 0.1 --top-k 50 --repeat-penalty 1.05

Add :Q4_K_M (or another quant tag) to the repo name to pick the file yourself.

The QAD Q4_0 file with the recommended sampling spelled out:

llama-cli -hf LiquidAI/LFM2.5-24B-A2B-GGUF \
    --hf-file LFM2.5-24B-A2B-QAD-Q4_0.gguf \
    --temp 0.1 --top-k 50 --repeat-penalty 1.05

Ollama:

ollama run hf.co/LiquidAI/LFM2.5-24B-A2B-GGUF:Q4_K_M

LM Studio: open lmstudio://open_from_hf?model=LiquidAI/LFM2.5-24B-A2B-GGUF or search the repo name in the model browser.

On a GPU box with full weights, vllm serve "LiquidAI/LFM2.5-24B-A2B" gives you an endpoint at http://localhost:8000/v1/chat/completions. Transformers needs transformers>=5.0.0.

Two settings worth checking before you blame the model:

1. Sampling. The card recommends temperature 0.1, top_k 50, repetition_penalty 1.05. The GGUF stores these as general.sampling.* metadata, so llama.cpp applies them automatically. Frontends that override the defaults with their own sampling (plenty do) will not. If output looks rambly, check this first.
2. Update your runtime. The architecture (lfm2moe) has been supported since LFM2-24B-A2B, but the tokenizer doubled to 128K. An old build that predates this release could trip on it. I haven't seen a report of that yet, but updating llama.cpp or Ollama costs you two minutes.

For agent harnesses (Hermes, OpenClaw, Pi), the card's example configs point them at a local endpoint, http://localhost:8080/v1, and the Hermes and OpenClaw ones set context length to 131,072.

Is it really fast? The 184 tok/s fine print

The card says the edge numbers were measured with llama.cpp, Q4_K_M and a 4K-token input context. That gives 184 tokens/s on an M5 Max. On AMD Ryzen AI Max+ 395 it's 114 tokens/s. Those are both top-end machines with lots of unified memory bandwidth. A 16 GB laptop spilling weights into slow RAM is a different story.

The bigger catch is that this is a reasoning-only model. Every answer starts with a chain of thought. A 2B-active MoE that writes 600 thinking tokens before replying "yes" can feel slower end to end than an 8B dense model that just answers. Liquid lists a DSpark draft model for speculative decoding as "coming soon", and someone is already asking about it in the discussions.

The LFM Open License and the $10M revenue limit

The model is under the LFM Open License v1.0. Commercial use is fine only if you are under the "Threshold", defined as annual revenue of $10,000,000 or more. Your "Legal Entity" includes affiliates that control you, are controlled by you, or share common control with you. So a small subsidiary of a big company doesn't get to count only its own revenue. Above the threshold, commercial use isn't licensed at all; you'd need a separate deal with Liquid.

For indie developers and most startups this is a non-issue. For anyone at a mid-size company, it means this isn't Apache 2.0, so don't treat it like Qwen when legal asks.

Should LFM2.5-24B-A2B replace your 8B model?

If your daily driver is an 8B dense model doing tool calls, JSON extraction or an agent loop, I think LFM2.5-24B-A2B is worth a weekend. The instruction-following and tool-use numbers are the strongest part of the card, and the active parameter count is under a third of what you're running now. If you mostly ask it coding or trivia questions, Liquid's own table says Qwen3.6-35B-A3B is the better pick, and an 8B won't embarrass itself either.

So, two questions for the comments. Will a 2B-active MoE that needs 14.5 GB of RAM actually replace your 8B dense daily driver, or is the memory cost a dealbreaker? And if you've loaded the GGUF: what tokens/sec are you getting, on what hardware, with which quant? Someone needs to check that 184.

Sources

• LiquidAI/LFM2.5-24B-A2B model card
• LiquidAI/LFM2.5-24B-A2B-GGUF
• Liquid AI launch blog, Oct 13, 2026
• LFM Open License v1.0
• Liquid AI model library
• GGUF repo discussions
• Ollama lfm2:24b (previous generation)