🔍 Read the full analysis: LFM2.5-24B-A2B Brings Fast, Private AI Agents On-Device on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Liquid AI has introduced LFM2.5-24B-A2B, a reasoning model designed for tool-using agents and local deployment. The company reports strong results on several agent and instruction-following benchmarks, but weaker scores on some coding, web and knowledge tests; the benchmarks and speed figures have not been independently verified in the supplied material.
Liquid AI has introduced LFM2.5-24B-A2B, a reasoning model designed for tool-using AI agents that can run on local hardware, as detailed in the original analysis. The company says the mixture-of-experts model has 24 billion total parameters but activates 2 billion per token, a design intended to lower the compute burden of repeated inference while supporting multi-step tasks.
Liquid AI describes LFM2.5 as a reasoning-only model that generates an explicit reasoning phase before its final answer and can make tool calls across multiple turns. The company says it trained the model within agent environments including OpenClaw, Hermes Agent, Pi and OpenCode. Its context window is 131,000 tokens, up from 32,000 in the preceding LFM2-24B-A2B model.
In company-reported results across 16 benchmarks, LFM2.5 scored 48.09 on LiveClaw, compared with 38.19 for Qwen3.6-35B-A3B, and 27.69 on AutomationBench, more than three times the next-highest listed score of 8.72. Liquid AI says it led or matched the compared models on five of eight agent and tool-use tests and led on all three instruction-following tests. These are the company’s evaluations, not independently replicated findings in the supplied material.
The results were mixed across other task types. The source says Qwen scored higher on WebArena Verified, GPQA Diamond and SWE-Bench Verified, among other tests; Gemma led on HLE, while Nemotron had the least negative score on AA-Omniscience. Liquid AI says it used vLLM and benchmark-recommended generation settings. Its reported decoding rates include 184 tokens per second on an Apple M5 Max and 114 on a Ryzen AI Max+ 395. It also reports more than 21,000 output tokens per second on one NVIDIA B300 at high concurrency.
Local Inference for Agent Workloads
Agents often make repeated model calls as they plan, use tools and respond to results. That can make compute requirements a practical barrier, particularly when a system is expected to operate on a user’s device or an organization’s own infrastructure. Liquid AI’s 2-billion-parameter active path is intended to reduce the amount of computation used per token compared with activating the full 24-billion-parameter model.
Local execution may let users keep prompts and outputs on hardware or infrastructure they control, depending on how the system is configured. The release is therefore relevant to developers weighing privacy, latency and operating costs alongside task performance. The available results do not establish that LFM2.5 will be faster or cheaper in every setup, or that it will outperform other models across general-purpose work.
The benchmark pattern points to a model aimed at workflows that rely on tool use and instruction following. Teams will need to test it against their own tasks, since the reported weaknesses in some web, coding and knowledge evaluations could matter more than agent benchmark gains for particular applications. Real-world costs also depend on memory use, power draw, quantization and the number of calls an agent makes—details not supplied for the reported speed figures.
As an affiliate, we earn on qualifying purchases.
What Changed From LFM2
LFM2.5 retains the architecture of LFM2-24B-A2B, which Liquid AI describes as a 40-layer model with 64 experts per mixture-of-experts block and top-four routing. The company says the design combines gated short convolutions with grouped-query attention. The most visible change for users is the longer context window, which grows from 32,000 to 131,000 tokens.
Liquid AI says pre-training expanded from about 17 trillion to 33 trillion tokens, followed by mid-training on reasoning, mathematics, code, instruction following, multilingual text and long-context material. The company also reports increasing its vocabulary from about 65,000 to 128,000 tokens to improve non-Latin tokenization. Post-training included supervised fine-tuning, teacher specialization, two stages of multi-domain on-policy distillation and agentic reinforcement learning.
The release is listed for use with inference systems including llama.cpp, MLX, vLLM, SGLang and ONNX, with a Transformers loading example also provided. That range of frameworks may make testing across different environments easier, but availability through an inference framework alone does not establish how much memory or processing capacity a particular deployment needs.
““Competitive with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks.””
— Liquid AI
As an affiliate, we earn on qualifying purchases.
Benchmark Claims Need Outside Testing
The benchmark and speed figures in the source are Liquid AI’s reported evaluations. The material does not provide independent replications, confidence intervals or detailed hardware configurations for each speed result. It also does not specify the exact quantization, memory footprint, power consumption or quality settings behind the figures, making direct comparisons difficult.
The reported rate of more than 21,000 output tokens per second is tied to high concurrency on a single NVIDIA B300; it should not be read as the expected rate for one user generating text on a local device. How the model performs across different agent harnesses, tool configurations and locally available hardware remains unclear.
The source material also does not state pricing, license terms or total workflow costs. Those details would help determine whether the model is practical for individual users, organizations or hosted services. It is also not yet clear how the benchmark performance translates to sustained agent runs, where latency and task completion can depend on the model, tools and surrounding software together.
As an affiliate, we earn on qualifying purchases.
Developers Test Real-World Use
The next useful evidence will come from developers testing LFM2.5 in their own agent setups and publishing comparable results for latency, memory, power use and task completion. Liquid AI says the model has day-one support across several inference frameworks, and its instructions describe serving it through an OpenAI-compatible local endpoint for agent harnesses such as Hermes Agent.
Independent benchmark runs and clearer deployment details, including model availability and licensing, would help show whether the reported advantages hold beyond the company’s tests. Until then, the release presents a model aimed at local agent workloads, with promising company-reported results on selected tasks and open questions about its costs and performance in everyday use.
high performance AI inference servers
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is LFM2.5-24B-A2B?
It is a reasoning model from Liquid AI designed for AI agents that use tools across multiple turns. The company describes it as a 24-billion-parameter mixture-of-experts model that activates 2 billion parameters per token.
Can the model run locally?
Liquid AI says it is designed for local hardware and lists support through frameworks including llama.cpp, MLX, vLLM, SGLang and ONNX. The supplied material does not specify the memory or hardware requirements for each setup.
Does it outperform other models?
Not across every test in the reported comparison. Liquid AI says LFM2.5 led or matched the group on several agent and instruction-following benchmarks, while other models scored higher on some web, coding and knowledge evaluations.
Are the speed and benchmark results independently verified?
The figures in the source are reported by Liquid AI. The supplied material does not include independent replications, confidence intervals or full hardware and configuration details for all results.
What information is still missing?
The source does not provide pricing, license terms, detailed memory requirements or power use. Independent tests across agent setups would also help establish how the model performs in practical deployments.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
