FULL STORY

Liquid AI LFM2.5: From Edge Agent to Multimodal

Liquid AI launched the open-source LFM2.5 model, enabling efficient on-device research agents. The lineup quickly expanded with the LFM2.5-VL-3B vision-language model, advancing lightweight multimodal capabilities.

2026-08-04 ~ 2026-08-13 · 5 episodes · 37 posts

Episode 1 · Liquid AI Releases 2.6B On-Device Agent Model LFM2.5 (2026-08-04, 18 posts)

Liquid AI released LFM2.5-2.6B, a new on-device agent model with only 2.69B parameters, capable of fully local multi-step planning, tool calling, and reasoning on phones, laptops, and robots, ensuring data never leaves the device and near-zero marginal cost per run. Official data shows 30 tok/s on mobile, and benchmarks surpass several larger models, quickly topping Hugging Face trending. Developer @nicodotdev built a fully autonomous research agent in-browser, validating the feasibility of complex agent tasks on-device with small models. This offers a cost-effective solution for mobile and offline AI applications.

Confirmed

  • Specs: 2.69B parameters, 34T training tokens, 128K context. Compatible with llama.cpp and GGUF at launch; MLX version optimized for Apple Silicon, supporting text generation and multilingual (Chinese, English, French, German, Arabic).
  • Core strengths: Dense mixture architecture, fully local multi-step planning and tool calling, zero marginal cost, suitable for simple high-throughput tasks. Official data via @BTALabs: 30 tok/s on mobile.
  • Benchmarks: Surpasses several larger models; topped Hugging Face trending after release.

Why it matters

  • On-device agent potential: Developer @nicodotdev built a fully in-browser research agent that plans, calls tools, checks results, and loops until completion, proving small models can handle complex agent tasks on-device.
  • Lower deployment barrier: Brings powerful tool-calling to edge devices, offering a cost-effective solution for mobile and offline AI.

Episode 2 · 2.6B Model Powers Browser-Based Research Agent, Open Source (2026-08-04, 5 posts)

Developer nicodotdev built a fully browser-local research agent on Liquid AI's newly released LFM2.5-2.6B model, now open-sourced with an online demo. The agent demonstrates a strong autonomous workflow, proving that a 2.6B-parameter small model can handle complex multi-step agent tasks on edge devices, not just text autocompletion.

Confirmed

  • Core model: Built on Liquid AI's LFM2.5-2.6B, runs entirely in the browser locally.
  • Workflow: After receiving a question, the agent autonomously creates a plan, calls tools, checks results, and loops until completion.
  • Advanced capabilities: It can ask clarifying questions, fetch local context, and delegate research to an isolated Wikipedia sub-agent to gather sources.
  • Output: Upon plan completion, it synthesizes evidence to generate cited answers and downloadable research reports.

Why it matters

  • Pushing small-model boundaries: Developer nicodotdev emphasizes this proves 2.6B-parameter models have sufficient planning, reasoning, and tool-use abilities, making edge-agent deployment a reality.

Episode 3 · Liquid AI 2.6B Model Matches DeepSeek in Tool Calling, 3.7x Faster (2026-08-06, 3 posts)

Liquid AI's 2.6B-parameter LFM2.5 model achieves 90 tokens/s on Apple M5 Pro and matches DeepSeek-V4 in tool calling while being 3.7x faster on RTX 5090s, showcasing strong edge potential.

Episode 4 · Liquid AI Releases Open-Source Edge Model LFM 2.5 (2026-08-09, 4 posts)

Liquid AI has released LFM 2.5, a 2.6-billion parameter open-source edge Agent model. Designed with an elegant architecture, it targets local deployment, allowing small devices to run independently without relying on the cloud.

Episode 5 · LiquidAI Releases LFM2.5-VL-3B Edge Multimodal Model (2026-08-12, 7 posts)

LiquidAI has officially released LFM2.5-VL-3B, a lightweight vision-language model designed for edge computing and on-device deployment. It outperforms several larger competitors across multiple benchmarks, demonstrating exceptional edge deployment efficiency and multimodal capabilities.

已确认

  • 要点：The model has a total of 3B parameters, built on the 2.6B parameter LFM2.5 language model backbone and combined with a SigLIP2 400M vision encoder (also referred to as SigLIP2 NaFlex in the post).
  • 要点：Pre-trained on approximately 34T tokens, it supports image-to-text generation and conversational capabilities.
  • 要点：The model focuses on cross-platform (mobile, web, desktop) capabilities, showing massive improvements in screen understanding, text and chart reading (OCR), physical world information extraction, object grounding, as well as rapid instruction execution and tool calling.
  • 要点：In various benchmark tests, the model's scores surpassed competitors that are 2.6 times its size.

为什么重要

  • 要点：Delivering flagship-level multimodal performance within a lightweight parameter constraint, this model provides an efficient multimodal interaction solution for mobile and edge devices, effectively lowering the hardware barrier for on-device AI deployment.