Deploy native brokers in every single place with LFM2.5-2.6B

LFM2.5-2.6B is constructed to energy succesful brokers solely on-device. It helps device calling and multi-step workflows whereas staying small and quick sufficient for on a regular basis {hardware}, from laptops to telephones. This allows builders to deploy brokers in every single place, preserve information personal on the gadget, and scale utilization with out a cloud inference invoice.

Finest-in-class agent: Aggressive with fashions 4x bigger on device use, instruction following, and multi-step agentic duties.
Agentic reinforcement studying: Educated inside the preferred agentic harnesses to enhance compatibility.
Environment friendly inference: 220 tok/s on an Apple M5 Max and 113 tok/s on an AMD Ryzen CPU, in underneath 2.5 GB of reminiscence.

How we constructed a dependable agentic mannequin for edge units

LFM2.5-2.6B is pre-trained on ~34T tokens, with a mid-training part that extends the context window to 128K. Put up-training then turns the bottom mannequin into an agent in 4 levels:

Supervised fine-tuning (SFT): two rounds of SFT, weighted closely towards agentic information like device use, net search, and harness trajectories.
Instructor specialization: prepare one specialist instructor per area (math, code, device use, and extra).
Multi-domain on-policy distillation (MOPD): distill the specialist lecturers right into a single pupil.
Agentic Reinforcement Studying (Agentic RL): run multi-turn RL inside actual agent harnesses, the place the mannequin learns to work throughout totally different instruments, system prompts, and multi-turn job environments.

The Agentic RL pipeline separates mannequin optimization, inference, and atmosphere execution into distinct parts. The Coaching Engine optimizes the mannequin, whereas the Rollout Engine generates actions utilizing the newest coverage. The RL framework orchestrates the coaching loop by launching rollouts, gathering trajectories and rewards, and updating the mannequin.

Actions are executed inside a Sandbox Service, the place the Blackbox Harness hosts the agent (e.g., OpenClaw or Hermes Agent) and coordinates interactions with the duty atmosphere. The Harness Proxy lets us deal with agentic harnesses as black containers with no modification, whereas transparently capturing the token-level trajectories wanted to reconstruct and validate RL coaching samples.

Benchmark outcomes

We evaluated LFM2.5-2.6B in opposition to fashions as much as ~4x its dimension on STEM, instruction following, device use, and agentic duties. It’s the smallest mannequin within the group, but it competes with and sometimes beats the remaining.

On your app, the strengths are instruction following and gear use. LFM2.5-2.6B tops each instruction-following benchmark right here, and each tool-use benchmark besides BFCLv4, the place solely the 9.7B Qwen edges forward. On agentic duties, it beats each Gemma fashions and stays even with the Qwens. It additionally leads on information and stays shut on math. Coding is the one place the bigger fashions preserve a transparent lead, so attain for one thing larger there.

Inference velocity on CPU and GPU

LFM2.5-2.6B ships with day-one assist throughout the inference ecosystem, together with llama.cpp, MLX, vLLM, SGLang, and ONNX.

CPU inference. Attributable to its environment friendly LFM2 structure, LFM2.5-2.6B is the quickest mannequin we examined, with decode speeds of 220 tokens/s on an M5 Max and 113 tokens/s on a Ryzen AI Max+ 395. At 30 tokens/s, it lets you run succesful brokers even on a cellphone.

GPU inference. LFM2.5-2.6B is the quickest mannequin in its dimension class, reaching nearly 15K output tokens per second at excessive concurrency, roughly 1.3B tokens per day on a single H100.

The way to use LFM2.5-2.6B

Attain for LFM2.5-2.6B whenever you want on-device brokers for high-volume workloads.

Set up the newest model of transformers (suitable with transformers>=5.0.0):

pip set up -U transformers

Then load and run the mannequin:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = “LiquidAI/LFM2.5-2.6B”
mannequin = AutoModelForCausalLM.from_pretrained(
model_id,
device_map=“auto”,
dtype=“bfloat16”,

)
tokenizer = AutoTokenizer.from_pretrained(model_id)

immediate = “What’s C. elegans?”
input_ids = tokenizer.apply_chat_template(
[{“role”: “user”, “content”: prompt}],
add_generation_prompt=True,
return_tensors=“pt”,
tokenize=True,
).to(mannequin.gadget)

output = mannequin.generate(
input_ids,
do_sample=True,
temperature=0.2,
top_k=80,
repetition_penalty=1.05,
max_new_tokens=512,
)
print(tokenizer.decode(output[0], skip_special_tokens=False))

LFM2.5-2.6B demo

Take a look at this browser demo of LFM2.5-2.6B powering a analysis agent. The agent helps you analysis particular questions and generates a abstract.

Get Began

Each LFM2.5-2.6B and LFM2.5-2.6B-Base can be found on Hugging Face right this moment.

With LFM2.5, we’re delivering on our imaginative and prescient of AI that runs anyplace. These fashions are:

Obtain: LFM2.5-2.6B-Base and LFM2.5-2.6B on Hugging Face.
Strive: run the WebGPU demo in your browser, no setup wanted.
Use in your harness: comply with our information on find out how to run an area agent, like OpenClaw, Hermes Agent, and Pi.

We will not wait to see what you construct.

Quotation

Please cite this text as:

Liquid AI, “LFM2.5-2.6B: Deploy Brokers All over the place”, Liquid AI Weblog, Aug 2026.

Or use the BibTeX quotation:

@article{liquidAI202626B,
creator = {Liquid AI},
title = {LFM2.5-2.6B: Deploy Brokers All over the place},
journal = {Liquid AI Weblog},
12 months = {2026},
observe = {www.liquid.ai/weblog/lfm2-5-2-6b},
}