LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents

LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents

At a Glance

Item Value
Repository LiquidAI/LFM2.5-24B-A2B
Family guide LFM2.5 guide (11 articles)
Publisher guide Liquid AI: models and licenses
Published 2026-10-12
License lfm1.0
Formats safetensors
Paper arXiv:2511.23404
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

LiquidAI has released “LFM2.5-24B-A2B", an inference model designed for fast and private agent operations on local devices. It adopts a Mixture of Experts (MoE) architecture that limits the number of active parameters operating during inference to approximately 2B (2.45B) out of a total parameter count of 24B. Built specifically for reasoning—generating a Chain of Thought before outputting responses—it is constructed to execute multi-step task processing and external tool calls at a low cost. Regarding the base architecture of this model, a technical report titled “LFM2 Technical Report" has been published, aiming to deliver high computational efficiency and performance under the latency and memory constraints of edge environments.

Specifications

Specifications listed in the published materials are as follows:

  • Total parameters: 24.0B
  • Active parameters: 2.45B
  • Architecture: MoE (Lfm2MoeForCausalLM, 40 layers consisting of 30 double-gated short convolution blocks and 10 GQA, 64 experts per MoE block, top-4 routing)
  • Context length: 131,072 tokens
  • Vocabulary size: 128,000

Performance

Comparison results with other models posted on the model card are as follows. All models were measured by the publisher in a vLLM environment using benchmark-recommended generation parameters.

→ Scroll horizontally to see all columns

Benchmark LFM2.5-24B-A2B Qwen3.6-35B-A3B Gemma-4-26B-A4B-it Nemotron-3.5-Lightning (30B-A3B)
AutomationBench 27.69 8.72 5.46 2.61
BFCLv4 65.01 65.00 57.09 53.78
ToolSandbox 80.71 75.02 72.09 75.13
LiveClaw 48.09 38.19 41.25 37.72
PinchBench 83.18 89.93 77.12 84.20
τ³-Bench Banking 8.04 7.63 4.54 7.01
WebArena Verified 26.75 37.62 25.62 28.69
GPQA Diamond 74.42 83.69 74.90 72.42
AA-Omniscience -24.90 -24.35 -59.39 -3.58
HLE 7.40 7.69 8.57 5.65
IFEval 96.75 93.06 92.00 94.19
IFBench 74.69 64.21 47.44 71.67
Multi-IF 84.38 81.53 82.19 82.34
MATH-500 96.04 96.73 95.66 96.40
SWE-Bench Verified (OpenCode) 51.00 66.20 25.80 53.40
SciCode 32.71 37.35 35.42 27.78

Measurements by the publisher show that this model records exceptionally high scores in tool utilization, instruction following, and agent processing. It marks 96.75 on IFEval, which measures accuracy in adhering to format specifications, representing the highest level among the comparison targets. Furthermore, it takes a commanding lead in AutomationBench with 27.69, more than triple the score of the runner-up Qwen3.6-35B-A3B (8.72), and stands alone at the top for BFCLv4 (65.01), ToolSandbox (80.71), LiveClaw (48.09), and τ³-Bench Banking (8.04). On the other hand, in knowledge-intensive evaluations such as GPQA Diamond (74.42), which tests PhD-level scientific knowledge, and HLE (7.40), a collection of extremely difficult questions created by experts in various fields, it falls short of models like Qwen3.6-35B-A3B, which have roughly 50% larger total and active parameters. It also stays at 51.00 on SWE-Bench Verified (OpenCode)—which measures the ability to fix real GitHub issues—trailing upper-tier models, and falls slightly short of other models with 96.04 on MATH-500, which measures high school to competition-level mathematics despite high accuracy.

Strengths and Use Cases

LFM2.5-24B-A2B is built specifically for autonomous AI agents operating in local environments and collaborative processing with external tools. The publisher recommends agent use cases where it performs reasoning, invokes tools, and runs stably on local hardware. Post-training includes supervised fine-tuning (SFT) and distillation from expert teacher models, alongside Agentic Reinforcement Learning within representative agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode, adapting to the tools, system prompts, and dialogue patterns of each harness.

Additionally, because it features a wide context length of 131,072 tokens, it is suited for multi-step long-term interactions and processing tasks requiring a long context. Supported languages include English as well as multilingual support for Japanese, Chinese, Korean, Arabic, French, German, Italian, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish, and others. The vocabulary size has also been expanded from 65,000 to 128,000 tokens, improving tokenization efficiency for languages containing non-Latin characters.

It is a reasoning-only model that inserts <think> tags to develop a thinking process before outputting responses, allowing it to execute tools after organizing complex procedures. As for the tool invocation procedure, when given JSON-format function definitions in the system prompt, it is designed to output Python-format function calls enclosed in <|tool_call_start|> and <|tool_call_end|> by default, receiving their execution results to generate the final response (it can also be instructed to output JSON-format function calls).

Note that the publisher does not recommend operating this model standalone for knowledge-intensive tasks. As shown in benchmark results, it falls short of large models in tasks requiring extensive knowledge, so a configuration that combines it with web search or external retrieval (RAG) to leverage its strength in tool-calling capabilities is recommended.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 24.0B parameters

Your VRAM Quantization File size Est. memory needed
80GB class (A100 / H100) F32 44.7GB 53.6GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): registered.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. The listed files are distributed in F32 (32-bit). Loading in a lower precision (BF16 or quantized) needs less memory, so treat this as the size of the files as distributed, not a minimum. This site has not measured the lower-precision figures. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — lfm1.0 (Commercial use allowed with conditions): LiquidAI’s LFM Open License v1.0. Commercial use is allowed, except by legal entities with annual revenue of $10 million or more (Section 5). Redistribution requires the license text and notice of changes.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Recent Models in the Same Size Class

Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
IFM/Jais-3-32B-GGUF 34.8B 24GB apache-2.0 Jais-3-32B-GGUF Arabic and English 32B Dense Model: 24GB+ VRAM (2026-11-17)
IFM/Jais-3-MoVA-36B-A4B-GGUF 37.4B 32GB apache-2.0 Jais-3-MoVA-36B-A4B-GGUF Text Generation Model: 32GB+ VRAM (2026-11-17)
tencent/Hy-Memory-30B-A3B 30.1B 12GB apache-2.0 Hy-Memory-30B-A3B Text Generation Model: 12GB+ VRAM, GGUF Builds (2026-11-17)
KRAFTON/Raon-1-mini-preview 28.8B 24GB apache-2.0 Raon-1-mini-preview Reasoning-Enabled Language Model: 24GB+ VRAM (2026-11-17)
ggml-org/Flint-GGUF 39.8B 32GB apache-2.0 Flint-GGUF Coding MoE Language Model: 32GB+ VRAM (2026-11-17)

How to Get It

LFM2.5-24B-A2B’s repository is available on Hugging Face and can be obtained directly without gated settings (prior agreement to terms of use). In addition to the original safetensors format, multiple formats and derived models are provided as follows:

  • Original format (safetensors): LiquidAI/LFM2.5-24B-A2B (for Transformers, vLLM, and SGLang)
  • FP8 format: LiquidAI/LFM2.5-24B-A2B-FP8 (for high-throughput GPU serving in vLLM and SGLang)
  • GGUF format: LiquidAI/LFM2.5-24B-A2B-GGUF (for llama.cpp and CPU environments)
  • ONNX Runtime format: LiquidAI/LFM2.5-24B-A2B-ONNX (for cross-platform use)
  • MLX format: LiquidAI/LFM2.5-24B-A2B-MLX-4bit (for Apple Silicon)

Additionally, a pretrained model LiquidAI/LFM2.5-24B-A2B-Base that can be used as a base for fine-tuning and other purposes is also published.

When running with Transformers, version transformers>=5.0.0 is required. You can load and run generation with the following procedure:

pip install -U transformers
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "LiquidAI/LFM2.5-24B-A2B"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype="bfloat16",
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

prompt = "What is C. elegans?"

input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
)["input_ids"].to(model.device)

output = model.generate(
    input_ids,
    do_sample=True,
    temperature=0.1,
    top_k=50,
    repetition_penalty=1.05,
    max_new_tokens=4096,
    streamer=streamer,
)

When integrating into an agent harness, serve it locally via OpenAI-compatible server features in llama.cpp, MLX, vLLM, SGLang, LM Studio, etc., and connect from each harness. For example, when using Hermes Agent, you can specify endpoints and context lengths with configuration commands like the following:

hermes config set model.provider custom
hermes config set model.base_url http://localhost:8080/v1
hermes config set model.default LFM2.5-24B-A2B
hermes config set model.context_length 131072
hermes config set model.api_mode chat_completions
hermes config set agent.tool_use_enforcement true

Related Articles

What to Read Next

Sources