LFM2.5-2.6B runs agents locally on laptops and phones with a 128K context window and tool calling — no cloud inference bill. LiquidAI built it as part of the LFM2.5 family, a line of hybrid models designed specifically for on-device deployment.
What you get:
- 2.6B parameter hybrid architecture
- 128K context window
- Tool calling and multi-step agent workflows
- OpenAI-compatible endpoint — drop-in for existing tooling
- Works with Hermes Agent, OpenClaw, and Pi out of the box
- Docker Model Runner support for local serving
Setup is two steps: serve the model behind an OpenAI-compatible endpoint, then point your agent harness at it. No fine-tuning, no adapter layer. If your harness already speaks OpenAI’s API, it already speaks LFM2.5.
NOTE
At 2.6B parameters, reasoning depth has real limits. Complex multi-hop planning will hit a ceiling faster than a 70B cloud model. The tradeoff is privacy and cost — not raw capability.
The value proposition is direct: private data stays on-device, usage scales without per-token costs, and latency drops by eliminating network round trips. For teams building agent workflows that touch sensitive data — health, finance, internal enterprise tools — that’s the actual draw, not the parameter count.
Available now on Hugging Face under the LiquidAI org. Free to download, no API key required.