Liquid AI Launches LFM2.5-2.6B: An On-Device Agentic Model Outperforming Larger Peers
helloiamleonie · x · 2026-08-04
Liquid AI has released LFM2.5-2.6B, a 2.6B parameter agentic model designed entirely for on-device operation. It can plan, call tools, and execute multi-step tasks locally on phones, laptops, PCs, and robots, ensuring zero data leaves the device and marginal inference costs remain near zero.
Key Specs & Benchmarks:
- Built on the LFM2.5 flagship hybrid architecture, pre-trained on 34T tokens with a 128K context length.
- Delivers comparable or better scores than models up to 4x its size, beating Qwen3.5-9B on ToolSandbox (77.83) and Gemma-4-E4B-it on Multi-IF (80.07).
- Early testers confirmed that the official benchmarks align perfectly with independent evaluations.
Related event: Liquid AI Launches LFM2.5-2.6B On-Device Agent Model(5 posts)→
More from coding & agent
- YC Open-Sources QM, a Multiplayer Agent Harness with Model Agnosticism — alex_verem · 2026-08-05
- Optimizing Codex Agents: A Workflow to Reduce Slop and Save Tokens — alexcovo_eth · 2026-08-05
- Using Codex for Data Science: Parallel Agents Tackle Housing, Schools, and Churn — gabrielchua · 2026-08-05
- Dynamic Model Router Matches Inference Levels Based on Context and KV Cache — testingcatalog · 2026-08-05
- Geek Workflow: Self-Made Phone App for Remote Voice Control of All tmux Sessions — yacineMTB · 2026-08-05
- Snorkel AI Builds Simulated Enterprise Environments to Train AI Agents on Complex Workflows — ajratner · 2026-08-05