The Experimental Feature That Makes Local Models Work

The Experimental Feature That Makes Local Models Work Better on Low-Power Hardware

OpenClaw 2026.5.21 added something small in the changelog that I'm betting a lot of people will find useful: experimental.localModelLean lets you enable lean local-model mode for one configured agent instead of globally.

This is relevant if you've been trying to run a local model on a machine that can't comfortably handle the full OpenClaw runtime plus the model inference.

What "Lean Mode" Actually Does

Lean mode reduces the resource footprint of the OpenClaw runtime when it's acting as the inference backend for a local model. Specifically:

Reduced context window overhead — the session system doesn't pre-load as much context for the agent
Lower memory baseline — the OpenClaw process itself uses less RAM when waiting for model responses
Simpler transcript handling — fewer transcript entries are kept in memory during the session

The result: a machine with 16GB of RAM that couldn't run a local model reliably with OpenClaw (because OpenClaw's runtime overhead was consuming too much memory alongside the model) can now do it.

The Config

To enable lean mode for a specific agent (not globally):

{
  "agents": {
    "list": [
      {
        "id": "local-model-agent",
        "experimental": {
          "localModelLean": true
        },
        "model": {
          "provider": "ollama",
          "name": "qwen3.5-8b"
        }
      }
    ]
  }
}

This is per-agent. You can have one agent running lean mode (for lighter workflows) and another running the full runtime (for heavier tasks) on the same machine.

Why This Matters for the Headless Use Case

If you're running OpenClaw on a headless server — a Raspberry Pi, a cheap VPS, an old laptop you repurposed as a home server — you probably want local model inference, not cloud API inference. But the resource overhead of the full OpenClaw runtime has been a barrier.

Lean mode reduces that barrier. It's not a magic solution — a 4GB model on a 2GB RAM machine is still going to struggle — but for the common case of "8B parameter model on 16GB RAM machine," lean mode makes it viable.

The xAI Device-Code OAuth Fix

The same release adds xAI device-code OAuth, which removes the localhost-browser callback requirement for remote servers. If you use xAI as your model provider and you've been stuck on setups where you can't authenticate because there's no browser on the server, this is the fix.

{
  "providers": {
    "xai": {
      "authMethod": "device-code"
    }
  }
}

This has been a friction point for headless xAI users. The fix is in 2026.5.21.

Lean mode: OpenClaw 2026.5.21, per-agent experimental flag. xAI device-code OAuth: same release, removes localhost callback requirement.

DEV Community