LFM2.5-2.6B Training Explained - SFT, Specialist Teachers, MOPD and Agentic RL

By Wei Jie Chee · 05 Aug 2026, 00:00 Z

Download printable cheat-sheet (CC-BY 4.0)
60-second takeaway
Liquid AI did not turn LFM2.5-2.6B into an agent with one broad fine-tuning run. It first trained a general instruction model, branched that checkpoint into domain specialists, distilled their capabilities back into one student on the student's own outputs, then trained the consolidated model inside real agent harnesses.
The result is a 2.69B-parameter, 128K-context model aimed at local tool use and multi-step workflows. Liquid reports inference under 2.5 GB and 220 tokens per second on an Apple M5 Max, but those are vendor measurements, not results we have independently reproduced.
The most interesting part is the training architecture: optimize skills separately, combine them with dense on-policy supervision, then test the combined policy where agents actually operate.

Liquid AI released LFM2.5-2.6B on 4 August 2026 as a compact text model for on-device agents. It has 2.69 billion parameters, a 131,072-token context window and open model weights under Liquid AI's LFM license.

The headline is local deployment. The deeper story is how Liquid tried to fit several competing capabilities into one small model without asking one training stage to solve everything at once.

The official pipeline has four post-training stages:

  1. two rounds of supervised fine-tuning;
  2. separate domain-specialist teachers;
  3. multi-domain on-policy distillation; and
  4. reinforcement learning inside agent harnesses.

A public Qwen chat summarized the design neatly as: train several small experts, distill them into one student on the student's own rollouts, then polish the result with agentic reinforcement learning.

That is a useful mental model. Liquid AI's release article supplies the details.

Start with the base model

Before post-training, LFM2.5-2.6B-Base was pretrained on about 34 trillion tokens. Liquid also extended the existing tokenizer to a 128,000-token vocabulary to improve support for non-Latin scripts, then added a dedicated mid-training phase to extend context to 128K.

The architecture contains 30 layers: 22 double-gated short-convolution blocks and eight grouped-query-attention blocks. This hybrid design is central to Liquid's efficiency pitch. Short convolutions handle much of the sequence processing, while attention layers retain the ability to connect information over longer distances.

Pretraining produces a capable next-token predictor, not yet a reliable assistant or agent. The four-stage post-training process is what shapes that base model into the released checkpoint.

Stage 1: two rounds of supervised fine-tuning

Supervised fine-tuning, or SFT, teaches the model from examples of desired inputs and outputs.