🔍 Read the full analysis: LFM2.5-24B-A2B For On-Device Agents: Speed Meets Privacy 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; independent validation and key deployment details are not provided.
Liquid AI has introduced LFM2.5-24B-A2B, a reasoning model designed to run tool-using agents on local hardware, as detailed in the original analysis. The company says its mixture-of-experts design has 24 billion total parameters but activates 2 billion per token, and reports competitive results on several agent and instruction-following benchmarks. The figures are company evaluations, not independently verified results.
Liquid AI describes LFM2.5 as a reasoning-only model that generates an explicit reasoning phase before its final answer and is built to make tool calls across multiple turns. The company says it trained the model within agent harnesses including OpenClaw, Hermes Agent, Pi and OpenCode. Its context window is 131,000 tokens, compared with 32,000 for the earlier LFM2-24B-A2B.
In a comparison across 16 benchmarks, Liquid AI reports that LFM2.5 scored 48.09 on LiveClaw, ahead of Qwen3.6-35B-A3B at 38.19. On AutomationBench, it scored 27.69; the next-highest listed score was 8.72. The company says the model led or matched the comparison group on five of eight agent and tool-use tests and led on all three instruction-following tests.
The results were not uniformly in LFM2.5’s favor. Liquid AI’s figures show Qwen scoring higher on WebArena Verified, GPQA Diamond and SWE-Bench Verified, among other tests; Gemma led on HLE, and Nemotron recorded 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. A separate rate above 21,000 output tokens per second was reported for one NVIDIA B300 at high concurrency.
Local Inference for Tool-Using Agents
Agent systems can make repeated model calls as they plan, use tools and respond to results. Each call adds compute demand, so a model that activates fewer parameters per token may be useful for developers trying to run such workflows on local machines or their own infrastructure. Liquid AI’s 2-billion active-parameter design is intended to reduce per-token inference cost while retaining a 24-billion-parameter model overall.
Local operation can also keep prompts and outputs on a user’s device or private infrastructure, depending on how the system is configured. That may appeal to developers handling sensitive material or seeking more control over deployment. It does not, by itself, establish that data stays local: users must check the full agent setup, including where tools run and whether any services make network requests.
The benchmark pattern points to a model worth testing for workflows centered on tool use and instruction following, not a general claim of superiority. Teams will need to judge its performance against their own tasks and account for latency, memory, power use, reliability and deployment constraints. The reported high-concurrency B300 throughput is not a typical single-user speed measure, and the supplied material does not quantify the resources needed to run the model locally.
As an affiliate, we earn on qualifying purchases.
How LFM2.5 Builds on LFM2
LFM2.5 retains 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. Its reported 131,000-token context window is a substantial increase from the predecessor’s 32,000 tokens.
Liquid AI says pre-training expanded from about 17 trillion to 33 trillion tokens, followed by mid-training focused on reasoning, mathematics, code, instruction following, multilingual text and long-context material. The company also says the vocabulary increased from about 65,000 to 128,000 tokens, with the aim of improving tokenization for non-Latin text. 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. Liquid AI describes serving it through a local OpenAI-compatible endpoint for agent harnesses. That framework support provides options for developers, but the supplied release information does not establish identical performance or hardware requirements across implementations.
““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.
Independent Tests and Deployment Costs
The benchmark and speed figures available here come from Liquid AI’s own evaluations. The source material provides no independent replication, confidence intervals or detailed hardware configurations. It does not specify the quantization, memory footprint, power consumption or quality settings behind each speed result. The B300 figure is explicitly for high concurrency, so it should not be compared directly with a single-user rate.
It is also unclear how performance varies across different agent harnesses, tool configurations and locally available devices. The release information provided does not state price, license terms or total workflow cost. Those details matter when deciding whether the model is practical for personal computers, team deployments or hosted infrastructure. No evidence in the supplied material establishes that local deployment automatically keeps all agent data on-device.
privacy-focused AI assistant devices
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Developer Tests Will Clarify Trade-Offs
The next useful evidence will come from developers testing LFM2.5 in their own agent setups and publishing results that can be compared on latency, memory use, power consumption and task completion. Independent replications using disclosed hardware and benchmark settings would help establish whether the company-reported advantages hold beyond its tests.
Developers considering the release will also need to check model availability, licensing and deployment instructions, then test the specific inference framework and agent configuration they plan to use. Until those details and broader results emerge, the benchmark scores are best treated as an initial company-reported evaluation, not a complete measure of real-world cost or capability.
high performance AI inference chips
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 agents that use tools. The company describes it as having 24 billion total parameters, with 2 billion active per token.
Can it run locally?
Liquid AI positions the model for local deployment and lists support through systems including llama.cpp, MLX, vLLM, SGLang and ONNX. The supplied material does not specify its memory requirements or performance across different local devices.
Does it outperform competing models?
Not across every test in the company’s comparison. Liquid AI reports strong results on some agent and instruction-following benchmarks, while competitors scored higher on several web, coding and knowledge evaluations. The results have not been independently verified in the supplied material.
How fast is the model?
Liquid AI reports 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; that figure is not a single-user generation rate.
What information is still missing?
The supplied release details do not provide independent benchmark results, full hardware and memory requirements, power figures, pricing or license terms. These details are needed to assess real-world performance and total deployment cost.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
