🔍 Read the full analysis: How To Run Fast, Private Agents On-Device With LFM2.5-24B-A2B on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Liquid AI has released LFM2.5-24B-A2B, a reasoning model designed for agents that use tools and can run on local hardware. The company reports strong results on several agent and instruction-following benchmarks, but its results are mixed across other tests and its speed claims have not been independently verified in the supplied material.
Liquid AI has introduced LFM2.5-24B-A2B, a reasoning model designed for tool-using agents that can run on local hardware. The company says the model has 24 billion total parameters but activates 2 billion per token, and reports leading or matching its comparison group on several agent and instruction-following benchmarks.
Liquid AI describes LFM2.5-24B-A2B as a reasoning-only model that generates an explicit reasoning phase before its final answer. It is designed to make tool calls across multiple turns, and the company says it was trained inside agent harnesses including OpenClaw, Hermes Agent, Pi and OpenCode. The model has a 131,000-token context window, up from 32,000 tokens in its predecessor, LFM2-24B-A2B.
In Liquid AI’s comparison across 16 benchmarks, LFM2.5 scored 48.09 on LiveClaw, versus 38.19 for Qwen3.6-35B-A3B, and 27.69 on AutomationBench, compared with the next-highest listed score of 8.72. The company 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 company-reported evaluations.
The results were not uniformly in LFM2.5’s favor. Qwen scored higher on WebArena Verified, GPQA Diamond and SWE-Bench Verified, among other tests; Gemma-4-26B-A4B-it led on HLE, while Nemotron-3.5-Lightning had the least negative score on AA-Omniscience. Liquid AI says it used vLLM and each benchmark’s recommended generation settings. It also reports decoding speeds of 184 tokens per second on an Apple M5 Max and 114 on a Ryzen AI Max+ 395, as well as more than 21,000 output tokens per second on one NVIDIA B300 at high concurrency.
Local Inference for Tool-Using Agents
Agents that reason through a task and call tools repeatedly can make many model requests, so the compute needed for each request affects whether a workflow is practical on local hardware. LFM2.5’s 2-billion-parameter active path is intended to lower per-token inference demands compared with activating all 24 billion parameters. That design makes the release relevant to developers weighing local deployment for agent workloads.
Local execution can keep prompts and outputs on a user’s machine or an organization’s infrastructure, depending on deployment and configuration. The benchmark pattern offers a reason to evaluate the model for tool use and instruction following, but does not establish broad superiority. Teams will need to compare it against alternatives on their own tasks, hardware and agent setups, including memory needs and sustained latency.
high performance local AI inference hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A Longer Context and Expanded Training
LFM2.5 keeps the architecture of LFM2-24B-A2B, which Liquid AI describes as a 40-layer mixture-of-experts model with 64 experts per block and top-four routing. The company says it combines gated short convolutions with grouped-query attention. The expanded context window is one of the clearest changes from the predecessor: 131,000 tokens compared with 32,000.
Liquid AI says pre-training grew from about 17 trillion to 33 trillion tokens, followed by mid-training in reasoning, math, code, instruction following, multilingual text and long-context material. The company also reports expanding the 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 model is listed for use with inference systems including llama.cpp, MLX, vLLM, SGLang and ONNX; the release material also provides a Transformers loading example. Liquid AI says it is competitive with Qwen3.6-35B-A3B on agentic and tool-use tasks despite Qwen having about 50% more total and active parameters. That comparison is the company’s characterization of its benchmark results.
““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 Settings Still Matter
The supplied material does not include independent benchmark replications, confidence intervals or detailed hardware configurations for the reported speed figures. It also does not specify the quantization, memory footprint, power use or quality settings behind each result. The more than 21,000-token-per-second figure is tied to high concurrency on a single B300 and should not be treated as a typical single-user generation rate.
It remains unclear how the model performs across different agent harnesses, tool configurations and local systems beyond those named. The material also does not state price or license terms, or provide a comparable estimate of the total cost of running a complete agent workflow. Those details will shape whether the release fits personal devices, team deployments or hosted services.
AI model deployment on personal computer
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Developers to Test Local Workflows
Liquid AI says the model has day-one support across several inference frameworks and describes serving it behind an OpenAI-compatible local endpoint for agent harnesses such as Hermes Agent. Developers can now test it in their own workflows and compare task completion, latency, memory use and power consumption against other models on the same hardware.
Independent results and fuller deployment details would help show whether the reported benchmark strengths hold in sustained use. Comparable tests should identify model settings, hardware, concurrency and tool configurations. Licensing and availability information will also matter for developers deciding whether to adopt the model.
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 tool-using agents. The company says it has 24 billion total parameters and activates 2 billion per token.Can the model run locally?
Liquid AI presents it as suitable for local hardware and lists support for systems including llama.cpp, MLX, vLLM, SGLang and ONNX. Actual performance will depend on the hardware and configuration.How did it perform on benchmarks?
Liquid AI reports strong results on several agent and instruction-following tests, including LiveClaw and AutomationBench. Its results were mixed across the full comparison, with competing models scoring higher on some web, coding and knowledge evaluations.Are the speed claims independently verified?
The supplied material presents the figures as Liquid AI’s evaluations and provides no independent replication or full hardware and configuration details.Primary source: Hugging Face · via ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
