OpenClaw v2026.7.2 Brings Native llama.cpp and Local Inference Support

OpenClaw v2026.7.2 introduces first-class local inference support, making it practical to run AI models entirely on your own hardware without relying on cloud API providers.

llama.cpp In-Process Inference

The release adds an in-process llama.cpp GGUF inference path that is RAM-gated — meaning the model is loaded and run directly within OpenClaw’s process, subject to available memory constraints. This enables entirely offline operation for supported GGUF-format models.

Guided Setup

The onboarding experience now includes local-provider detection that automatically identifies available local inference backends during setup. Users can:

  • Select from detected local providers
  • Choose a strongest-model option from locally available models
  • Download models directly from the web or macOS setup flow
  • Enable lean mode for resource-constrained environments

Baseten Model API

In addition to llama.cpp, OpenClaw adds Baseten Model API support, enabling connection to self-hosted models deployed on Baseten’s infrastructure.

Provider Catalog Discovery

Models are now dynamically discovered from live provider catalogs, so available model lists stay current without requiring a restart or manual refresh.

See the OpenClaw releases page for the full changelog.

← Back to News