Liquid AI released LFM2.5-2.6B, a 2.69 billion parameter model built for on-device deployment with a 128,000 token context window, according to the model's Hugging Face listing.
The model scored 80.07% on the Multi-IF instruction-following benchmark, compared with 55.67% for Alibaba's 4 billion parameter Qwen3.5-4B, and scored 77.83% on the ToolSandbox benchmark for tool use, versus 75.55% for the same Qwen model, according to Liquid AI's listing.
Running quantized, the model reaches 220 tokens per second on an Apple M5 Max chip and 113 tokens per second on an AMD Ryzen processor, while using under 2.5 gigabytes of memory, the listing says.
Liquid AI recommends the model for agentic workloads, tool use, data extraction, retrieval-augmented generation and long-context tasks, and says it is not intended for coding or knowledge-intensive work. The weights are released under the company's LFM1.0 license.
For builders, this is a model sized to run alongside an app instead of behind an API call. The benchmark gap against a model nearly twice its size is the actual pitch: a lower capability ceiling, but a real shot at fitting an agent's tool-calling loop on the device already in a user's hand.