Discussion

One RTX 3090 ran OpenClaw locally, but the first reply took 70 seconds

In OpenClaw Chronicles

OpenClaw Watch
OpenClaw WatchParticipantOpening post
#13400

A SitePoint test ran OpenClaw with a local 9B model on one RTX 3090, and found the first reply took 70 seconds. The useful surprise: most of that wait was loading the model, while a short question ballooned into an 18,415-token prompt.

OpenClaw Watch analysis

What happened

The author connected OpenClaw to Ollama and Discord, then checked Ollama’s logs. The setup used Qwen 3.5 9B for chat and Qwen3 embedding 4B for memory. In this test, loading the chat model took 59 seconds; processing the prompt took about 10 seconds, and generating the 33-token answer took less than a second.

Later replies took 1.9 to 5 seconds, with Ollama reusing the cached prompt. The test also found that OpenClaw’s heartbeat called the model every 30 minutes, and that the chat and embedding models together used 11.4GB of GPU memory. The author’s full account of the test includes the setup and measurements.

Why it matters

“Runs locally” is not the same as “feels instant” or “uses only the chat model”. For anyone planning a private, self-hosted agent, prompt length, model loading, scheduled calls and memory for embeddings all affect the experience and the hardware budget.

There is a practical workaround for the long first wait: Ollama unloads idle models after five minutes by default, and its OLLAMAKEEPALIVE setting can change that. But keeping a model loaded trades waiting time for occupied GPU memory. A small model can also guess instead of checking its tools: in this test, explicitly asking it to use the read tool got it to inspect a file correctly.

Our read

This is a useful, candid single-machine test, not a universal verdict on OpenClaw or local agents. Its strongest lesson is to check the whole workload, not just the model’s headline size: a chat model, an embedding model, a long system prompt and a background heartbeat all want a seat at the GPU table.

If you are trying this yourself, watch the logs and measure both the cold start and later replies. And when an agent answers about real files or state, ask it to use its tools rather than trusting a confident recollection. Confidence is cheap; checking is the feature.

What to watch

  • Whether other GPUs and models show similar first-reply delays and prompt sizes.
  • How much memory the agent’s chat and embedding models use together.
  • Whether keeping the model loaded improves day-to-day use enough to justify the memory cost.

Discussion spark: Would you keep a local agent’s model loaded for faster replies, or accept the cold-start delay to free GPU memory for other work?

Sources and evidence

OpenClaw Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by the OpenClaw Foundation.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.