Liquid AI releases LFM2.5-24B-A2B, a 24B model with 2B active parameters for on-device agents

Liquid AI says LFM2.5-24B-A2B is competitive with Qwen3.6-35B-A3B on five of eight agentic and tool-use benchmarks it reported, scoring 48.09 on LiveClaw against 38.19 and 27.69 on AutomationBench against 8.72. Qwen3.6-35B-A3B leads on GPQA Diamond, HLE-style knowledge tasks, SWE-Bench Verified and several other benchmarks, according to the same tables.

The company reports 184 tokens per second decoding on an Apple M5 Max and up to 21K output tokens per second on a single NVIDIA B300 GPU under sustained concurrent load. The model keeps 40 layers and a context window extended from 32,768 to 131,072 tokens, and its vocabulary doubled from 65K to 128K tokens. Liquid AI also describes a post-training pipeline that trains the model inside agent harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.

The model is available through llama.cpp, MLX, vLLM, SGLang and ONNX. Liquid Console API access is priced at $0.10 per million input tokens and $1.00 per million output tokens.

Developers who want an agent to run continuously on local hardware now have an open-weight option that Liquid AI says handles tool calls well, though its own tables show it trailing larger models on knowledge-heavy and coding tasks.

Liquid AI recommends pairing the model with retrieval or web search for knowledge-heavy workloads, so teams should plan for that in their architecture rather than relying on the model's built-in knowledge.