🔍 Read the full analysis: Run AI Agents Locally With LFM2.5-24B-A2B For Speed And 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. Company-reported tests show strong results on some agent and instruction-following benchmarks, but weaker results on several knowledge, web and coding tests; the speed figures have not been independently verified in the supplied material.
Liquid AI has introduced LFM2.5-24B-A2B, a reasoning model designed for agents that call tools and can run on local hardware. The company reports gains on several agent and instruction-following benchmarks, while its own results also show that competing models lead on some web, coding and knowledge tests.
The model uses a mixture-of-experts design with 24 billion total parameters and 2 billion active per token, according to Liquid AI. The company says this design is intended to lower the compute needed for each inference while supporting multi-step reasoning and tool calls across multiple turns. It describes the model as reasoning-only, with an explicit reasoning phase before its final answer.
LFM2.5 has a reported 131,000-token context window, up from 32,000 tokens in its predecessor, LFM2-24B-A2B. Liquid AI says it trained the model inside agent harnesses including OpenClaw, Hermes Agent, Pi and OpenCode. The company lists support for inference systems such as llama.cpp, MLX, vLLM, SGLang and ONNX, and provides a Transformers loading example.
In Liquid AI’s comparison across 16 benchmarks, LFM2.5 scored 48.09 on LiveClaw, ahead of Qwen3.6-35B-A3B at 38.19. Its AutomationBench score was 27.69, compared with 8.72 for the next-highest listed model. Liquid AI says LFM2.5 led or matched the comparison group on five of eight agent and tool-use tests and led on all three instruction-following tests. These are company-reported evaluations, not independent results in the supplied material.
The comparison was not uniformly favorable to LFM2.5. Qwen scored higher on WebArena Verified, GPQA Diamond and SWE-Bench Verified, among other tests; Gemma led on HLE, and Nemotron had the least negative result 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 Apple M5 Max and 114 on Ryzen AI Max+ 395. A separate figure of more than 21,000 output tokens per second refers to high concurrency on one NVIDIA B300.
Local Inference for Agent Work
Agents that reason and repeatedly call tools can make many model requests during a single task. A model with 2 billion active parameters per token could reduce the compute required for each request compared with a model that activates all of its parameters, though the supplied figures do not establish the actual memory, energy or total workflow cost.
Running a model locally can keep prompts and outputs on a user’s machine or organization’s infrastructure, depending on deployment and configuration. That can matter for developers handling sensitive material or seeking more control over latency and hosting. The release’s reported benchmark pattern gives teams a reason to evaluate LFM2.5 for tool use and instruction following, but does not show that it will be the best choice for general-purpose work.
As an affiliate, we earn on qualifying purchases.
How LFM2.5 Extends Its Predecessor
LFM2.5 keeps 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 model combines gated short convolutions with grouped-query attention. Its stated changes include a larger context window and expanded training.
Liquid AI says pre-training grew from about 17 trillion to 33 trillion tokens, followed by mid-training on reasoning, math, code, instruction following, multilingual text and long-context material. The company also reports increasing 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.
“Competitive with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks.”
— Liquid AI
AI model deployment on personal computer
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Benchmark and Deployment Limits
The supplied material does not provide independent replications or confidence intervals for the benchmark scores. It also does not specify the exact quantization, memory footprint, power use or quality settings behind each speed result. The 21,000-plus token-per-second figure is tied to high concurrency on a single B300, so it does not describe a typical single-user rate.
Performance may vary with agent harnesses, tool configurations and local hardware beyond the systems named. The material does not give license terms, pricing or a comparative total-cost measure for running a complete agent workflow. Those details will affect who can use the model and under what conditions.
large language model for local use
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Local Runs
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 compare its behavior in their own setups, including task completion, latency and resource use.
The next useful evidence will include independent benchmark replications with full settings, alongside measurements of memory, power and sustained performance. Licensing, availability and deployment costs also remain relevant questions. Until those details and comparable real-world results are available, the company’s reported scores offer a starting point for evaluation rather than a complete picture of how LFM2.5 performs in routine agent work.
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, with 2 billion active per token.
Can it run locally?
Liquid AI lists support for inference systems including llama.cpp, MLX, vLLM, SGLang and ONNX. Actual local performance will depend on the hardware and configuration.
How does it compare with other models?
Liquid AI reports strong results on some agent and instruction-following tests. Its own comparison also shows other models scoring higher on several web, coding and knowledge benchmarks.
Are the speed figures independently verified?
The supplied material attributes the figures to Liquid AI’s evaluations. It does not provide independent replication or enough configuration detail to judge how the results translate to other systems.
Primary source: Hugging Face · via ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
