Skip to content

LFM2.5-24B-A2B: A Practical Mac Guide

October 14, 2026 · 5 min read

LFM2.5-24B-A2B on Mac is a new option for running a reasoning and tool-use model locally. Liquid AI released the open-weight model on October 13, 2026, with first-party GGUF and MLX versions available immediately. Its mixture-of-experts design activates only about 2.45 billion of its 24 billion parameters for each token, but you still need enough memory to load the full model.

In short: start with the 13.6 GB MLX 4-bit or 14.5 GB GGUF Q4_K_M version on a Mac with at least 24 GB of unified memory. A 16 GB Mac is too tight for a comfortable agent workload once the operating system, runtime, and conversation context are included. These are practical estimates from published file sizes, not Enclave AI performance tests.

What Is LFM2.5-24B-A2B?

LFM2.5-24B-A2B is the largest and final model in Liquid AI’s LFM2.5 family. It is a reasoning-only model: the chat template starts a reasoning trace before the final answer rather than offering a switch between thinking and direct response.

Liquid expanded the previous generation’s context window from 32,768 to 131,072 tokens, doubled the vocabulary to 128K tokens, and trained the model for multi-turn tool use and agent workflows.

That makes the model interesting for local coding assistants, research workflows, and personal automations. It does not make every agent private by default: a locally running model can still call web search, cloud tools, analytics, or remote APIs.

Why Do 24B Total and 2.45B Active Parameters Matter?

The name describes a mixture-of-experts model:

  • 24B total parameters represent the model’s full set of weights.
  • About 2.45B active parameters per token do most of the computation for each generation step.
  • 64 experts per MoE block with top-4 routing let the model select a small subset of experts for each token.

The smaller active count can improve generation speed, but it does not turn the model into a 2.45B download. All experts remain available for routing, so the full 24B weights must fit in storage and working memory.

Quantization is therefore the practical bridge to consumer Macs. If the format names are unfamiliar, our LLM quantization explainer covers Q4, Q5, and Q8 in plain language.

How Much Mac Memory Does LFM2.5-24B-A2B Need?

Liquid AI’s first-party files provide a useful starting point. The GGUF repository lists Q4_0 at 13.6 GB, Q4_K_M at 14.5 GB, Q5_K_M at 17 GB, Q6_K at 19.7 GB, Q8_0 at 25.5 GB, and BF16 at 48 GB. The MLX exports range from 13.57 GB at 4-bit to 47.95 GB at BF16.

Mac unified memory Practical starting point Why
16 GB Choose a smaller model Even the smallest first-party export leaves almost no room for macOS, the runtime, and context.
24 GB 4-bit MLX or Q4_K_M GGUF The weights fit with some headroom; begin with shorter contexts and one workload.
32 GB Q4_K_M, Q5_K_M, or 6-bit MLX More headroom for context and agent tools, though the exact limit depends on the runtime.
48 GB Q6_K or Q8_0 BF16 nearly fills memory with weights alone, so a quantized build is more practical.
64 GB or more Q8_0 or BF16 Higher precision fits more comfortably, but BF16 still needs additional runtime and context memory.

This table is a fit guide, not a speed benchmark. Model-file size is only the floor: the KV cache grows with the conversation, and the inference engine needs additional memory. Our KV cache guide explains why a 14 GB download can require meaningfully more than 14 GB at runtime.

How Do You Run LFM2.5-24B-A2B Locally?

The official GGUF card provides a one-command path through llama.cpp:

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

That downloads the Q4_K_M build and starts a local OpenAI-compatible server for compatible clients and agent harnesses. Confirm that any tools you connect do not silently send prompts elsewhere.

Apple Silicon users can also use the first-party MLX exports. Liquid’s model card says to use mlx-lm 0.32.0 or later; earlier versions can load the model without an error but route its experts incorrectly, producing different output. The 4-bit export is the sensible first test for most 24 GB and 32 GB Macs.

Liquid also publishes a Q4_0 GGUF made with quantization-aware distillation. The company says it was trained to recover accuracy lost during Q4_0 compression. That is different from proving it is better for your tasks, so compare it with Q4_K_M on a representative prompt set before standardizing on either file.

What Do the Benchmarks Actually Show?

Liquid reports strong agent and tool-use results, but the numbers are vendor-run. In its release table, LFM2.5-24B-A2B leads or matches the compared models on five of eight agentic benchmarks and all three instruction-following benchmarks. The same table shows larger competitors ahead on several knowledge and coding evaluations, including GPQA Diamond, SWE-Bench Verified, and SciCode.

The company also reports 184 generated tokens per second on an Apple M5 Max and says the model ran up to 1.7 times faster than the compared MoE models on-device. Do not project that figure onto every Mac: chip generation, memory bandwidth, quantization, prompt length, thermal state, and runtime all affect real performance. We did not independently reproduce Liquid AI’s results.

A better evaluation for a personal agent is task-specific:

  1. Build prompts that resemble your real work.
  2. Check tool selection, argument accuracy, final answers, and failure recovery separately.
  3. Run the same set with the exact quantization you intend to keep.
  4. Measure memory, prompt-processing time, and generation speed on your Mac.
  5. Require confirmation before destructive, financial, or external actions.

Is a Local LFM2.5 Agent Private?

Local inference can keep prompts and model computation on your Mac instead of sending every message to a hosted model.

The agent’s tools define the wider boundary. Web searches, email connectors, cloud storage, remote logs, and crash reporting can still transmit data. Review each integration, keep sensitive arguments out of logs, and use narrow permissions. Our guide to why offline AI matters covers the difference between local computation and a fully private application.

The model is released under the LFM Open License v1.0, not Apache 2.0 or MIT. Read the current license before redistribution or commercial deployment.

The Bottom Line

LFM2.5-24B-A2B is a credible new local-agent model for Macs with enough unified memory.

For most people, the practical entry point is a 4-bit build on a 24 GB or 32 GB Mac. Treat Liquid’s benchmarks as useful vendor evidence, not a guarantee; test the exact quantization, context length, agent tools, and safety boundaries that your workflow needs.

Sources