TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Liquid AI has released LFM2.5-24B-A2B, a mixture-of-experts reasoning model designed to run agent workflows on local hardware. The company reports strong results on several tool-use and instruction-following benchmarks, but its own tests show larger models leading on some knowledge and coding tasks.
Liquid AI has announced LFM2.5-24B-A2B, a 24-billion-parameter reasoning model designed for agents that run on users’ own hardware. The company says the model activates 2 billion parameters per token, supports tool calls and multi-step tasks, and is intended to provide local inference rather than requiring every interaction to be sent to a remote service.
Liquid AI describes the release as a reasoning-only model: it generates an explicit reasoning trace before its final answer, according to the company. It is built for use inside agent frameworks and can be served locally through an OpenAI-compatible backend. The report lists support across Transformers, llama.cpp, MLX, vLLM, SGLang and ONNX, though the source does not provide independent confirmation of performance across those implementations.
The company evaluated the model against three other mixture-of-experts systems on 16 benchmarks covering agent tasks, knowledge, instruction following, math and coding. On LiveClaw, Liquid AI reports a score of 48.09, compared with 38.19 for Qwen3.6-35B-A3B. It also reports 27.69 on AutomationBench, more than three times the next-highest listed score of 8.72, and 65.01 on BFCLv4, narrowly above Qwen’s 65.00.
Results were not uniformly higher. LFM2.5 scored below some competitors on WebArena Verified, GPQA Diamond, HLE and coding tests, including SWE-Bench Verified with OpenCode, where it scored 51.00 against Qwen’s 66.20. Liquid AI says the models were tested using vLLM and each benchmark’s recommended generation settings. Those figures are company-reported evaluations, not an independent assessment in the supplied material.
Local Agents Without Remote Inference
The release targets developers who want agent workflows to run on local hardware, potentially keeping prompts and task data on the device or within an organization’s own infrastructure. Local deployment can also reduce dependence on a hosted model endpoint. However, running a model locally does not by itself guarantee privacy: data handling depends on the surrounding software, configuration and any external tools the agent uses.
Liquid AI’s central efficiency claim is that only 2B of the model’s 24B parameters are active per token. That design is intended to make repeated reasoning and tool-use steps less costly than activating the full parameter count each time. The company reports decoding at 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, and says performance reaches more than 21,000 output tokens per second on one NVIDIA B300 at high concurrency. The B300 figure is a high-concurrency GPU result, not a directly comparable measure of a single user’s response speed.
For developers, the benchmark mix suggests a potential fit for tool use and instruction-following tasks, rather than a general advantage across all workloads. Liquid AI itself advises pairing the model with retrieval or web search for knowledge-heavy applications. That caveat matters for agents expected to answer factual questions or act on current information: a model’s ability to call tools does not establish that it will choose the right tool or verify its results reliably in every setting.
As an affiliate, we earn on qualifying purchases.
Architecture and Training Changes
LFM2.5-24B-A2B builds on Liquid AI’s LFM2-24B-A2B and retains its reported architecture: 40 layers, 64 experts per mixture-of-experts block, and top-four routing, alongside gated short convolutions and grouped-query attention. The company says the context window has expanded from 32,000 to 131,000 tokens. It also reports expanding the tokenizer vocabulary from about 65,000 to 128,000 tokens, with the goal of improving tokenization for non-Latin text.
Liquid AI says pre-training increased from roughly 17 trillion to 33 trillion tokens. Mid-training added 2 trillion tokens at a 32,000-token context length, weighted toward reasoning, mathematics, coding, instruction following and multilingual text, followed by 300 billion tokens at 131,000 context, weighted toward long-context data. Tool calls and multi-turn trajectories were included in both phases, according to the report.
Post-training used five stages, the company says: supervised fine-tuning, teacher specialization, two rounds of multi-domain on-policy distillation, and agentic reinforcement learning inside harnesses including OpenClaw, Hermes Agent, Pi and OpenCode. The model is available under the identifier LiquidAI/LFM2.5-24B-A2B; the report gives a Transformers version requirement of 5.0.0 or later. The supplied source does not state a release date or provide licensing details.
“LFM2.5-24B-A2B is a 24B-parameter reasoning model with just 2B active parameters, built for fast, private agents on-device.”
— Liquid AI, in its Hugging Face report
As an affiliate, we earn on qualifying purchases.
Independent Tests and Privacy Details
The supplied report does not include independent benchmark results, full hardware and configuration details for every comparison, or evidence that the scores will transfer directly to different prompts and agent setups. It also does not define the exact concurrency level behind the B300 throughput figure. Results can vary with quantization, software versions, memory limits and the way an agent is configured.
“Private” describes the intended on-device use, but the source does not provide a security audit, data-retention analysis or guarantees about the complete agent stack. An agent may send information to external services when it uses web search or other remote tools. The model’s licensing terms, hardware requirements for common local deployments, and any safety or reliability evaluation beyond the listed benchmarks are not specified in the material provided.
AI model deployment on personal computer
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing in Local Agent Setups
Developers can test the model through supported inference software and connect a local serving endpoint to an agent harness. Liquid AI’s report gives an example using Hermes Agent and an OpenAI-compatible local backend, with the context length set to 131,072 tokens. The next practical test is whether the model’s reported benchmark strengths hold for real tasks on the hardware and tools a developer plans to use.
Further evaluation would help establish how performance changes across quantized versions, different devices and independent agent configurations. Readers should also check the model’s current repository for licensing, installation requirements and updated implementation support, details not included in the supplied report. Until third-party tests and deployment-specific measurements are available, the reported scores and speeds should be treated as Liquid AI’s results rather than universal performance guarantees.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is LFM2.5-24B-A2B?
It is Liquid AI’s 24-billion-parameter mixture-of-experts reasoning model, designed for agent tasks and reported to activate 2 billion parameters per token.
Does it run entirely on a device?
Liquid AI says it is designed for local inference and lists support in several inference frameworks. Whether a complete agent stays on-device depends on the hardware, configuration and whether the agent calls remote tools or services.
How did it compare with other models?
In Liquid AI’s tests, it led or matched competitors on several listed tool-use and instruction-following benchmarks, including a 48.09 LiveClaw score. Other models scored higher on some knowledge and coding benchmarks. These are company-reported results.
What hardware speeds did Liquid AI report?
The company reports 184 output tokens per second on an Apple M5 Max, 114 on a Ryzen AI Max+ 395, and more than 21,000 output tokens per second on a single NVIDIA B300 at high concurrency. The B300 result reflects a different, high-concurrency setup.
Is its privacy claim independently verified?
Not in the supplied report. Local operation can limit data sent to a hosted model, but Liquid AI does not provide a security audit or a guarantee that an entire agent workflow avoids external services.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
