TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Liquid AI has introduced LFM2.5-24B-A2B, a reasoning model designed for tool-using agents that can run on local hardware. Its published benchmarks show strengths in several agent and instruction-following tasks, while larger rivals score higher in some knowledge and coding tests; the speed figures come from Liquid AI’s own evaluations.

Liquid AI has introduced LFM2.5-24B-A2B, a reasoning model built for agents that use tools and run on local hardware. The company says the mixture-of-experts model has 24 billion total parameters but activates 2 billion per token, a design intended to reduce inference cost while supporting multi-step tasks.

Liquid AI describes LFM2.5-24B-A2B as a reasoning-only model: it produces an explicit reasoning phase before its final answer and is designed to make tool calls across multiple turns. The company says it has been 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.

The company reports results across 16 benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning. LFM2.5 scored 48.09 on LiveClaw, compared with 38.19 for Qwen, and 27.69 on AutomationBench, more than three times the next-highest listed score of 8.72. It led or matched the comparison group on five of eight agent and tool-use tests and led on all three instruction-following tests, according to Liquid AI.

Results were not uniformly stronger. 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 the models were evaluated using 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, plus more than 21,000 output tokens per second on one NVIDIA B300 at high concurrency.

Local Agents With Lower Per-Token Compute

The release targets a practical constraint for agent builders: running a model that reasons and repeatedly calls tools can require substantial compute. By activating 2 billion parameters per token, rather than all 24 billion, LFM2.5 is designed to make frequent local inference more economical. Local deployment can also keep prompts and outputs on a user’s own machine or infrastructure, depending on how the system is configured.

The benchmark pattern suggests the model may suit workflows centered on tool use and following instructions. It does not establish that it will outperform alternatives across general-purpose tasks. Liquid AI’s own results show gaps in some web, coding and knowledge evaluations, so teams choosing a model will need to weigh task-specific performance alongside speed, memory use and deployment requirements.

A Larger Context for LFM2

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. It also retains a hybrid design combining gated short convolutions with grouped-query attention. The company says training expanded from about 17 trillion to 33 trillion pre-training tokens, followed by mid-training focused on reasoning, math, code, instruction following, multilingual text and long-context material.

Liquid AI says 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 model is listed for use through several inference systems, including llama.cpp, MLX, vLLM, SGLang and ONNX; the source also gives a Transformers loading example.

““Competitive with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks.””

— Liquid AI

Independent Results and Real-World Costs

The benchmark and speed figures in the supplied material are Liquid AI’s reported evaluations; no independent replication, confidence intervals or detailed hardware configurations are provided there. The source does not specify the exact quantization, memory footprint, power use or quality settings behind each speed result. Its figure of more than 21,000 output tokens per second is explicitly tied to high concurrency on a single B300, so it should not be read as a typical single-user generation rate.

It is also unclear how performance will hold up across agent harnesses, tool configurations and locally available hardware beyond the systems named. The release material does not give a price, license terms, or a comparative measure of total cost for a complete agent workflow. Those details affect whether the model is practical for personal devices, teams or hosted deployments.

Deployment and Independent Testing

The next useful evidence will come from developers running the model 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 behind an OpenAI-compatible local endpoint for agent harnesses such as Hermes Agent.

Readers evaluating the release should look for full benchmark settings, independent replications and deployment details, including model availability and licensing. Those would help clarify how the reported advantages translate from company tests to sustained, real-world agent use.

Where I land

My assessment is that LFM2.5-24B-A2B is a focused local-agent release with a credible case to test, especially for tool use and instruction following. Its reported LiveClaw and AutomationBench scores stand out, and a 2-billion-parameter active path could make repeated inference easier to fit on edge hardware. Those results are promising signals, not proof that it will be the best choice for a particular agent.

The strongest counterargument is that the model trails competitors on several demanding web, coding and knowledge benchmarks, while the speed claims come from the vendor and depend on specific hardware and concurrency. I would raise my assessment if independent evaluations confirmed the agent-task gains and documented memory, power and end-to-end latency on widely available devices. I would lower it if those tests showed that local constraints or tool errors erase the benchmark advantage.

Key Questions

What is LFM2.5-24B-A2B?

It is a 24-billion-parameter mixture-of-experts reasoning model from Liquid AI. The company says 2 billion parameters are active for each token.

Does it run on personal hardware?

Liquid AI positions it for local use and reports speeds on an Apple M5 Max and a Ryzen AI Max+ 395. Actual performance depends on the hardware, model configuration and inference software.

Which tasks does it perform well on?

In Liquid AI’s evaluations, it led or matched the comparison models on five of eight agent and tool-use benchmarks and led on the three instruction-following benchmarks. Rivals scored higher on several knowledge and coding tests.

Are the reported speeds independently verified?

The supplied results are attributed to Liquid AI. The material does not include independent replication or all configuration details needed to compare the speed figures across deployments.

Source: Hugging Face