Breaking

Liquid AI Releases LFM2.5-2.6B, a 2.6B On-Device Agent Model

August 31, 2026 at 11:48 EDT

Foundation Models
AI Agents
Open Source

Liquid AI has released LFM2.5-2.6B, a compact agentic language model built to plan, call tools, and complete multi-step tasks entirely on phones, laptops, PCs, and robots, with no cloud inference in the loop. The post-trained checkpoint and a companion base model, LFM2.5-2.6B-Base, were published on August 4, 2026.

Aug 4, 2026 · Liquid AI

A 2.6B agent that never leaves the device

LFM2.5-2.6B plans, calls tools, and works through multi-step tasks on phones, laptops, PCs, and robots, with no cloud API in the loop.

2.69B

Parameters

~34T

Training tokens

128K

Context window

220

tok/s on M5 Max

Same job, much smaller stack

Vendor tests put the 2.6B model against Qwen3.5-9B at 9.7B parameters — nearly 4x the size — on tool use.

2.6B

LFM2.5-2.6B

9.7B

Qwen3.5-9B

ToolSandbox: 77.83 for LFM2.5-2.6B vs 76.44 for the 9.7B rival.

How a 2.6B model became an agent

1 · SFT

Two fine-tunes on tool use and traces

→

2 · Teachers

Domain experts, then distillation

→

3 · Agentic RL

Live harnesses: Hermes, OpenClaw, Pi

Where it lands

Leads instruction-following scores in the comparison set, stays under 2.5 GB, and holds about 30 tok/s on a phone.

Where it does not

Not recommended for agentic coding or knowledge-heavy work. Long tool chains can stall or re-call the same tool.

The product boundary

Use it on-device when a cloud model is not an option: privacy, latency, or a meter that never turns off.

Free commercial weights below $10M annual revenue. Larger firms need a paid license.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.