LFM2.5-24B-A2B launches: A local agent model with 2B active parameters

Liquid AI introduced its LFM2.5-24B-A2B reasoning model in an announcement published on Hugging Face on 13 October 2026. The company positions it for agents that can run on-device, call tools, and handle multi-step tasks. The model has 24 billion parameters in total, with 2 billion activated for each token processed.

A focus on longer context and reasoning

LFM2.5-24B-A2B retains the architecture of the previous LFM2-24B-A2B model. According to Liquid AI, the design uses 40 layers, 64 experts in each MoE block, and top-4 routing. The model was developed as a version that follows an explicit reasoning process before responding.

The context window has grown from 32K tokens in the previous version to 131K tokens. The vocabulary has also expanded from 65K to 128K tokens; the company says this is intended to improve token usage in languages that do not use the Latin alphabet. Pretraining has scaled from approximately 17 trillion tokens to approximately 33 trillion. The company also describes additional training stages focused on long context, reasoning, mathematics, code, and multilingual text.

Tool use stands out in benchmarks

Liquid AI says it compared the model across 16 benchmarks with Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning. In the company’s results, LFM2.5-24B-A2B scored 27.69 on AutomationBench, 65.01 on BFCLv4, 80.71 on ToolSandbox, and 48.09 on LiveClaw. The comparison score for Qwen3.6-35B-A3B on LiveClaw was 38.19.

According to the announcement, the model matched or outperformed its rivals on five of eight tests for tool use and agent workflows; it also led on all three instruction-following benchmarks. The company notes, however, that larger models retain an advantage on knowledge-intensive tests such as GPQA Diamond and the coding-focused SWE-Bench Verified. For knowledge-heavy tasks, Liquid AI therefore recommends pairing the model with search or retrieval systems. These results are the company’s own evaluations; the source states that all models were tested with vLLM.

Local deployment and hardware results

The model is available with support for llama.cpp, MLX, vLLM, SGLang, and ONNX. According to Liquid AI’s measurements, it generates 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395. The company also reports more than 21,000 output tokens per second on a single NVIDIA B300 GPU under high concurrency. These figures are manufacturer-reported results; the same speeds are not guaranteed on different hardware or settings.

The model and LFM2.5-24B-A2B-Base are available as open weights on Hugging Face. The announcement does not include pricing. It says an OpenAI-compatible backend can be set up on a local server and connected to supported agent systems. In the example setup, the context length is set to 131,072; the port depends on the server software used.

Implications for architecture and visualization workflows

For architecture firms, interior designers, and visualization teams, the model’s appeal lies in the ability to run a tool-calling agent on their own hardware. This could be worth exploring for teams looking to use local agents for repetitive text and information-processing tasks; however, the announcement does not say that the model integrates directly with specific CAD, BIM, or visualization software. Any potential use should therefore be considered within the limits of supported agent infrastructure and the tool connections teams set up.

Before deciding on local deployment, check hardware performance, setup compatibility, and the terms governing the open weights. The announcement gives speed results for different CPUs and GPUs, but does not specify the required memory or hardware costs in Turkey. Teams should therefore test the model on their own tasks and compare the results with their existing workflows.