Liquid AI Launches LFM2.5-2.6B: An On-Device Agentic Model
JosephJacks_ · x · 2026-08-06
Liquid AI has released the LFM2.5-2.6B model, focusing on agentic capabilities that run entirely on-device. It can plan, call tools, and execute multi-step tasks on phones, laptops, and robots, ensuring data privacy and near-zero marginal costs.
Key Specs:
- Pre-trained on approximately 34T tokens.
- Features the flagship LFM2.5 hybrid architecture.
- Context length and vocabulary size are both 128K.
- Balanced intelligence per watt, customizable on a single GPU.
Performance:
The model achieves comparable or better scores than models nearly 4x its size. For instance, it scores 77.83 on ToolSandbox (beating Qwen3.5-9B at 76.44) and 80.07 on Multi-IF (outperforming Gemma-4-E4B-it at 77.35).
Related event: Liquid AI Launches LFM2.5-2.6B Edge Agent Model(17 posts)→
More from coding & agent
- Building AI Agent Memory: A Framework for Skills and Notes CRUD — donotfire · 2026-08-06
- shadcn Reflects on AI Agents: When Copying Code Becomes Free, What Happens to Creativity? — shadcn · 2026-08-06
- OpenClaw's Task Flow: Solving Runtime Challenges for Durable, Long-Lived Agents — RichardsonDx · 2026-08-06
- RacterMX Launches MCP Server: Let AI Manage Email Infrastructure via Natural Language — Objective-Test-5374 · 2026-08-06
- Tutorial: Implementing Entra ID Authentication and Authorization for MCP Servers — datawiza · 2026-08-06
- Benchmarking Fallback Models for Agents: Why Failure Visibility Beats Raw Quality — AccomplishedLab3697 · 2026-08-06