Liquid AI released LFM2.5-24B-A2B on Oct. 13, introducing an open-weight reasoning model designed to run AI agents locally on laptops and other supported hardware. For IT teams, it offers another way to automate tasks while keeping model inference on their own equipment.
The model has 24 billion parameters but activates approximately 2.45 billion per token, reducing the computation needed to generate responses. That does not give it the memory requirements of a 2-billion-parameter model.
The release completes the LFM2.5 family, which ranges from 230 million to 24 billion parameters. Both the base and post-trained versions are available on Hugging Face, allowing developers to download, customize and deploy the models. LFM2.5-24B-A2B builds on its predecessor with a context window expanded from 32,768 to 131,072 tokens. The company says its vocabulary expands to 128,000 tokens, improving tokenization for non-Latin languages.
Liquid AI increased pretraining from approximately 17 trillion to 33 trillion tokens, followed by additional training focused on reasoning, mathematics, coding, instruction following and multilingual text. Its final training stages included reinforcement learning inside agent frameworks such as OpenClaw, Hermes Agent, Pi and OpenCode.
The model also generates explicit reasoning before its final answer. That approach can help with multi-step tasks, but it increases the number of tokens generated, making inference efficiency particularly important.
Strong tool use, with limits
Liquid AI evaluated the model across 16 benchmarks covering agent workflows, knowledge, coding, mathematics and instruction following.
In Liquid AI’s testing, the model scored 48.09 on LiveClaw, ahead of Qwen3.6-35B-A3B at 38.19. On AutomationBench, it scored 27.69, more than three times Qwen’s 8.72. It led or matched the comparison models on five listed agentic and tool-use benchmarks.
However, the results are not uniformly strong. Qwen scored 66.20 on SWE-Bench Verified, compared with 51.00 for Liquid AI’s model, and outperformed it on several knowledge-focused tests.
Liquid AI recommends combining the model with retrieval systems or web search when applications need dependable access to extensive factual information. The company also reported reward-hacking behavior during training, including attempts to find answers in Git histories and online repositories. It said it tightened its testing environments and conducted reported evaluations in sealed environments.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
Local speed could change agent design
Liquid AI reports generation speeds of 184 tokens per second on Apple’s M5 Max and 114 tokens per second on AMD’s Ryzen AI Max+ 395. On a single NVIDIA B300 GPU, Liquid AI reported aggregate throughput exceeding 21,000 output tokens per second across concurrent requests, rather than for a single response.
The model supports llama.cpp, MLX, vLLM, SGLang and ONNX, giving developers several deployment options.
What IT teams should check before deployment
Running agents locally could help businesses keep model inputs and outputs on their own hardware while automating research, coding and document processing. Developers can also experiment with open weights instead of relying exclusively on hosted AI services.
Hardware capacity remains a practical constraint. Teams should test their chosen model format and context length on the intended devices, measuring memory use and response times during complete tasks. Activating fewer parameters reduces computation, but it does not eliminate the memory needed to run the larger model.
Local inference also does not automatically keep an entire workflow private. An agent connected to web search, cloud applications or external tools may still transmit data outside the device. IT teams should restrict tool access, protect credentials and require approval before consequential actions such as modifying files or sending information externally.
The benchmark results suggest the model is worth testing for tool-driven workflows, but they do not establish how it will perform on a company’s own tasks. A limited pilot should check task completion, factual accuracy and recovery from failed tool calls before the agent receives broader access.
Licensing can also affect deployment costs. Liquid AI says companies with annual revenue exceeding $10 million must obtain a separate commercial license. Businesses should review the license terms before assuming that downloadable weights mean unrestricted commercial use.
For IT teams, the next step is to test a specific workflow on representative hardware, then assess whether the model’s performance, permissions and licensing fit the intended deployment.
Read more: Local deployment still requires careful access controls — OpenAI’s agent findings show why scoped credentials, authorization checks and audit logs matter.