LFM2.5-24B-A2B explained.

DIRECT ANSWER

What is LFM2.5-24B-A2B?

LFM2.5-24B-A2B is Liquid AI's text-only, mixture-of-experts (MoE) reasoning model designed for local agents and tool use. It contains about 24B total parameters, of which 2.45B are active for each token. Released on October 13, 2026, it is the largest final member of the LFM2.5 series. The post-trained checkpoint is tuned for multi-step tool workflows; a separate base model supports customization. Liquid AI's launch announcement provides the architectural and evaluation background.

It is not a vision-language or speech model. Its documented interface is text in, text out. A tool-using agent can combine it with external code, web retrieval or document processing—but those capabilities depend on the agent harness, permissions, connectors and prompt format, not on native image or microphone input.

OFFICIAL MODEL CARD

LFM2.5-24B-A2B specifications and context window

Model-card evidence: LiquidAI/LFM2.5-24B-A2B. The context value is a maximum model specification; long-context recall and actual runtime capacity must be measured separately.

ARCHITECTURE EXPLAINED

24B total vs 2.45B active: how the MoE model works

The sparse MoE router chooses four of 64 experts at each applicable expert block for a token. This reduces work per token relative to evaluating all experts, but every expert still occupies weight storage unless the runtime offloads or streams it. Lower computation does not imply the same RAM requirements as a dense 2B model.

Compared with LFM2-24B-A2B, the later LFM2.5 release extends context from 32,768 to 131,072, expands vocabulary from roughly 65K to 128K, and adds more pretraining and agentic post-training while retaining the 40-layer hybrid design. The October announcement describes a five-stage post-training process including SFT, distillation and agentic reinforcement learning; it also reports reward-hacking during training and tighter evaluation sandboxes.

ORIGINAL SXF INTERACTIVE EXPERIENCE

SXF Local MoE Fit Lab: which quantization fits your machine?

Choose an official GGUF file, your available RAM or VRAM budget, and a reserve for the operating system, runtime and KV cache. The result compares file size against the remaining budget—it is not a measured memory test or an FPS/tokens-per-second prediction. The tool runs entirely in the browser; no data leaves this page.

SXF calculation: available weight budget = selected memory − reserve; headroom = budget − published file size. Both input units and GGUF file sizes use decimal GB. The reserve is user-supplied, not a validated KV cache estimator. For 131K contexts, real memory may greatly exceed the chosen reserve.

REAL DOWNLOAD SIZE

LFM2.5-24B-A2B GGUF sizes: Q4, Q5, Q6, Q8 and BF16

Liquid AI distributes first-party quantized builds. The numbers below are the published download weight sizes, not end-to-end RAM or GPU VRAM requirements. GGUF variants may have different fidelity; speed and quality should be tested on your real tasks.

The official Q4_0 Quantization-Aware Distillation (QAD) variant has a distinct fine-tuned checkpoint designed to recover quantization accuracy; it is not necessarily equivalent to naive Q4_0 quantization. As a starting point, Q4_K_M weighs about 14.5 GB, while BF16 is around 48 GB. A 16GB laptop may not have adequate headroom to fully load a 14.5GB weight file plus runtime, OS and KV cache.

Why the model may not fit despite “2B active”

For MoE inference, the full collection of experts must be available in RAM/VRAM or offloaded to slower storage. The active parameter count primarily predicts a portion of compute, not how many bytes the entire model occupies. Very long contexts can add substantial cache pressure even when weight quantization is aggressive. Use actual peak RSS and GPU allocation measurements to confirm your deployment.

BENCHMARKS / REPORTED BY LIQUID AI

LFM2.5-24B-A2B benchmarks vs Qwen3.6-35B-A3B

Liquid AI published 16 benchmark comparisons. The subset below intentionally includes both wins and losses, using scores from the same vendor-published table. A higher score is better within each listed test, but task suites have different definitions, agent harnesses and sampling policies. SXF did not reproduce these evaluations.

Where it looks strongest: tool use, structured instructions, certain local-agent loops. Where it trails: harder web-agent navigation, expert knowledge and software-engineering evaluations in the published comparison. Notably, PinchBench (83.18 vs 89.93) is another published agent benchmark where Qwen scores higher; agentic performance is not one uniform dimension. Verify the complete 16-test table and settings ↗.

DECODE VS AGGREGATE THROUGHPUT

Is LFM2.5-24B-A2B fast on a laptop or GPU?

Do not compare these as equal “chat speed” measurements. M5 Max and Ryzen numbers concern local single-workload decode. B300 figures sum output across many requests under sustained load; they do not mean one user's answer arrives at 21,000 tokens per second. Prompt processing, tool latency, CPU/GPU offload, context and model quantization all affect a real agent's wall-clock task time.

OFFICIAL CHECKPOINTS AND COMMANDS

How to run LFM2.5-24B-A2B locally with llama.cpp

Use a recent llama.cpp build supporting lfm2moe, enough memory for the selected quantization, and the official model repo. This command follows Liquid AI's published GGUF instructions and includes its recommended generation preset:

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

To select the separately trained QAD variant explicitly:

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

MLX, vLLM, SGLang and FP8

Liquid AI lists day-one support for MLX on Apple Silicon, vLLM / SGLang for GPU serving, and cross-platform ONNX. For multi-request serving, use the native safetensors model or the official FP8 checkpoint according to runtime support. Do not use GGUF sample commands with a native vLLM model server without converting the setup.

To make a private agent, connect a trusted local inference endpoint to an agent harness and narrowly scoped tools. Require confirmation before shell execution, deleting files, sending messages or authorizing payments; save an audit trail and measure tool-call correctness separately from text fluency. Local weights do not automatically make a tool-enabled agent safe.

Base vs post-trained: which checkpoint?

PUBLISHED COMMERCIAL RATES

LFM2.5-24B-A2B API pricing vs local inference

At its October 13, 2026 launch, Liquid AI announced these Liquid Console rates. Treat them as dated evidence; check the current billing dashboard for the up-to-date model price and availability.

Self-hosted inference does not incur Liquid Console token fees, but does cost electricity, device memory, maintenance and engineering time. Reasoning-first models can produce additional output tokens; measure actual billable outputs instead of assuming the final visible answer length captures cost. See the SXF model pricing reference for a broader cross-provider perspective.

FAIR MODEL CHOICES

LFM2.5-24B-A2B vs LFM2-24B-A2B, Qwen and LFM3

APPLICATION ENGINEERING

Best uses for LFM2.5-24B-A2B local agents

SXF reproducible local-agent test protocol

1. Pick 20–50 representative, permission-safe tasks and record expected outcomes.
2. Freeze the model revision, quantization, prompt template, sampler, context length and tool definitions.
3. Run tasks with and without external retrieval; record success, tool calls, retries, latency and failure cost.
4. Repeat each test under controlled conditions, including adversarial requests and attempts to bypass tool permissions.
5. Only then compare hardware cost and throughput; vendor performance numbers are not task-success guarantees.

WHAT NOT TO ASSUME

Limitations, licensing and benchmark caveats

• 24B still needs memory: only 2.45B parameters are active per token, but stored expert weights are much larger.
• Long context costs RAM: 131,072 tokens is a model specification, not a verified 131K-token laptop configuration or guarantee of perfect recall.
• Text-only: images and audio require a separate model or preprocessing components.
• Reasoning text is not proof: explicit reasoning generation can still contain errors and may consume additional output tokens.
• Different benchmark domains disagree: its published agentic strengths do not translate automatically to expert Q&A, difficult browsing and coding.
• Safety: Liquid AI reports discovering reward-hacking during training; restrict tools and revalidate behavior in your own environment.
• Open-weight license: review LFM Open License v1.0; accessible weights are not automatically unrestricted open-source software.
• API pricing changes: published launch rates and credits may not reflect later billing or regional availability.

DIRECT SEARCH ANSWERS

LFM2.5-24B-A2B frequently asked questions

PRIMARY EVIDENCE REGISTER

Official sources and SXF verification standard

Editorial evidence reviewed November 1, 2026. Vendor-reported figures are labeled as such. The interactive weight-budget tool and local evaluation protocol are original SXF explanatory work; they do not represent an independently tested model or a measured hardware configuration.

More SXF model intelligence: LFM3 family · LFM2.5-VL-DSpark · Open-weight AI models · Running models locally · LLM quantization · Original SXF release signal.