🔍 Read the full analysis: A Practical Guide To Fast, Private On-Device Agents With LFM2.5-24B-A2B 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 that can run on local hardware. The company reports strong scores on several agent and instruction-following benchmarks, but results vary by task, and the published speed figures have not been independently verified in the supplied material.
Liquid AI has introduced LFM2.5-24B-A2B, a reasoning model built for tool-using agents and local deployment. The company says its mixture-of-experts design contains 24 billion total parameters but activates 2 billion per token, with a 131,000-token context window; its benchmark results show strengths in several agent tasks, alongside weaker scores on some knowledge, web and coding evaluations.
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 over multiple turns. The company says it trained the model inside agent harnesses including OpenClaw, Hermes Agent, Pi and OpenCode. Its context window is 131,000 tokens, compared with 32,000 for its predecessor, LFM2-24B-A2B.
In company-reported results across 16 benchmarks, LFM2.5 scored 48.09 on LiveClaw, versus 38.19 for Qwen3.6-35B-A3B, and 27.69 on AutomationBench, more than three times the next-highest listed result of 8.72. Liquid AI 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 varied by task. Qwen scored higher on WebArena Verified, GPQA Diamond and SWE-Bench Verified; 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 evaluated the models using vLLM and the generation settings recommended for each benchmark.
Lower Compute for Local Agents
Agents that reason through a task and call tools repeatedly can make many model requests, so the compute needed for each token can affect whether a workflow is practical on local hardware. LFM2.5’s 2-billion-parameter active path is intended to reduce per-token inference cost compared with activating all 24 billion parameters. That design goal may matter to developers building agents that run on personal devices or their own infrastructure.
Local deployment can keep prompts and outputs within a user’s machine or organization, depending on system configuration. The announcement does not establish that local processing is automatically private: data handling also depends on the surrounding software, connected tools and deployment choices. The reported benchmark pattern makes the model a candidate for agent and instruction-following workloads, but does not show that it is the best choice for general-purpose reasoning or coding.
Liquid AI also reports decoding speeds of 184 tokens per second on an Apple M5 Max and 114 on a Ryzen AI Max+ 395. It reports more than 21,000 output tokens per second on a single NVIDIA B300 at high concurrency. These company figures may help developers decide what to test, but the high-concurrency result does not describe typical single-user generation speed.
high performance local AI inference hardware
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From LFM2 to LFM2.5
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 it combines gated short convolutions with grouped-query attention. The main stated changes include the longer context window, expanded training and additional post-training focused on reasoning and agent use.
According to Liquid AI, pre-training expanded from about 17 trillion to 33 trillion tokens. It says later training covered reasoning, mathematics, code, instruction following, multilingual text and long-context material, and that the vocabulary grew 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 company lists support through inference systems including llama.cpp, MLX, vLLM, SGLang and ONNX, and provides a Transformers loading example. Liquid AI characterizes the model as competitive with Qwen3.6-35B-A3B on agentic and tool-use tasks despite Qwen’s larger parameter count. That comparison is the company’s characterization of its own results.
““Competitive with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks.””
— Liquid AI
private on-device AI agent software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmark and Deployment Gaps
The supplied release material gives company-reported benchmark and speed figures; it does not include independent replication, confidence intervals or detailed hardware configurations for each speed result. Exact quantization, memory use, power consumption and quality settings behind the speed tests are also not specified. The more than 21,000-token-per-second figure is tied to high concurrency on one B300 and should not be treated as a single-user rate.
It remains unclear how results will vary across agent harnesses, tool configurations and other locally available hardware. The material also does not provide a price, license terms or a comparative estimate of the total cost of running a complete agent workflow. Those details would affect the practical trade-offs for individuals, teams and hosted deployments.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Availability
Developers can test the model through supported inference frameworks and report how it performs in their own agent setups. Comparisons that publish task completion, latency, memory use and power consumption alongside benchmark settings would help show how the release performs outside Liquid AI’s evaluations.
Further details on licensing, model availability, quantization and hardware requirements would help teams assess deployment costs. Until those are available, the company’s benchmark and speed figures provide an initial basis for evaluation, rather than independently confirmed evidence of performance across real-world workloads.
AI development tools for local deployment
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 tool-using agents. The company says it has 24 billion total parameters and activates 2 billion per token.
Can it run locally?
Liquid AI designed it for local hardware and lists inference support in systems including llama.cpp, MLX, vLLM, SGLang and ONNX. Actual hardware requirements and performance will depend on the deployment configuration.
Does it outperform other models?
Not across every reported task. Liquid AI’s results show advantages on several agent and instruction-following tests, while competing models scored higher on some web, coding and knowledge benchmarks.
Are the reported speed figures independently verified?
The supplied material presents them as Liquid AI’s evaluations and provides no independent replication. It also does not specify all hardware and configuration details behind the figures.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
