A compact AI model with agent support that runs on smartphones, 'LFM2.5-2.6B,' has been released, allowing AI agents to be run in a local environment.
LFM2.5-2.6B: Deploy Agents Everywhere — Blog — Liquid AI
https://www.liquid.ai/blog/lfm2-5-2-6b
Unlike agents that rely on cloud APIs, local agents can perform inference for free, have lower latency, and offer superior privacy protection. Furthermore, according to Liquid AI, the elimination of per-token costs allows agents to be parallelized at scale on local hardware, enabling them to run background tasks that consume millions of tokens without marginal cost.
The following graph compares the benchmark results of LFM2.5-2.6B with those of 'Gemma 4 E28 it (5.18B)', 'Gemma 4 E48 it (8B)', 'Qwen3.5 (4.7B)', and 'Qwen3.5-9B (9.7B)'.
LFM2.5-2.6B is a 2.6B parameter model specifically trained for agent workloads, pre-trained with approximately 34T tokens of training data. A dedicated 128K context expansion phase is included during training to enable the model to handle the long inputs required for agent workloads.
The following diagram summarizes the four-stage post-training pipeline for fine-tuning the base model LFM2.5-2.6B-Base to LFM2.5-2.6B. Specifically, the processes used are supervised fine-tuning, supervised learning, multi-domain on-policy distillation, and agent-based reinforcement learning.
According to Liquid AI, LFM2.5-2.6B is the fastest model they've tested in terms of prompt loading and response generation, thanks to its efficient LFM2 architecture. It decodes 220 tokens per second on the Apple M5 Max and 113 tokens per second on the Ryzen AI Max+ 395, while keeping memory usage under 2.5GB. It can process up to 30 tokens per second on smartphones, meaning you can run a high-performance agent instantly and privately on your own device.
The following graphs compare the processing power per unit time for each model, benchmarked in an SGLang 0.5.16 environment with 1024 input tokens, a maximum of 256 output tokens, and BF16 settings, with an average of 3 runs for each parallelism level. The LFM2.5-2.6B is the fastest model in its size class, generating approximately 15,000 output tokens per second at high parallelism levels, and is said to be able to process approximately 1.3 billion tokens per day on a single NVIDIA AI GPU 'H100'.
The following is a demo of LFM2.5-2.6B actually running within the Liquid Agent harness on a smartphone. Task planning, tool calls, and actual task execution are all completed on the device without making any cloud API calls.
On-Device Agents Running 24/7: LFM2.5-2.6B Plans a Wedding Trip - YouTube
LFM2.5-2.6B-Base and LFM2.5-2.6B can be downloaded from Hugging Face.
LiquidAI/LFM2.5-2.6B-Base · Hugging Face
https://huggingface.co/LiquidAI/LFM2.5-2.6B-Base
LiquidAI/LFM2.5-2.6B · Hugging Face
https://huggingface.co/LiquidAI/LFM2.5-2.6B
Related Posts:
