---
title: "Run OpenClaw with a Local LLM — Agent-ready Models"
canonical: "https://build.nvidia.com/rtx/openclaw/agent-ready-models.md"
---

# Agent-ready models

Agent-ready models are tuned for **agentic workloads** — tool calling, reasoning traces, and long multi-turn sessions. Use this tab to pick a recommended local model for your hardware platform, then return to the **Instructions** tab to connect OpenClaw to the OpenAI-compatible endpoint.

## Recommendations by hardware platform

| Hardware platform | Recommended model | Example handle / tag | Inference path |
| ----------------- | ----------------- | -------------------- | -------------- |
| **DGX Spark** | Agent-ready Qwen3.6-35B-A3B (NVFP4) | `nvidia/Qwen3.6-35B-A3B-NVFP4` | vLLM — see [Serve LLMs with vLLM → Agent-ready Models](https://build.nvidia.com/playbooks/vllm/agent-ready-models) |
| **RTX** (24GB+ VRAM) | Qwen3.6 27B | `qwen/qwen3.6-27b` (LM Studio) or `qwen3.6:27b` (Ollama) | LM Studio, Ollama, or vLLM |
| **RTX** (12–16GB VRAM) | Qwen 3.5 9B / Gemma 4 12B | backend-specific tags | LM Studio or Ollama |
| **RTX** (6–8GB VRAM) | Qwen 3.5 4B | backend-specific tags | LM Studio or Ollama |

Only platforms listed in the Supported hardware platforms table (Overview) are listed above.

## Context window guidance

For OpenClaw, set the local server context window to **at least 32K tokens**. If your hardware platform has additional memory headroom, **64K or higher** is recommended so multi-turn agent sessions and skills have room to work.

## Large-memory Linux hardware platforms — vLLM agent-ready path

On large-memory Linux hardware platforms, use the agent-ready `nvidia/Qwen3.6-35B-A3B-NVFP4` recipe from the Serve LLMs with vLLM playbook’s [Agent-ready Models](https://build.nvidia.com/playbooks/vllm/agent-ready-models) tab. That path serves an OpenAI-compatible API (typically at `http://localhost:8000/v1`) suitable for OpenClaw.

1. Follow the Serve LLMs with vLLM playbook’s [Agent-ready Models](https://build.nvidia.com/playbooks/vllm/agent-ready-models) tab for the launch command for your hardware platform.
2. Verify:

```bash
curl http://localhost:8000/v1/models
```

3. In OpenClaw, set provider `baseUrl` to `http://localhost:8000/v1` and set `id` / `name` to `nvidia/Qwen3.6-35B-A3B-NVFP4`. Match `contextWindow` to the server `--max-model-len` (the agent-ready recipe uses a large context suitable for multi-turn agent sessions).

## Discrete-GPU hardware platforms — LM Studio / Ollama / vLLM

On discrete-GPU hardware platforms, choose a model that fits available VRAM (table above), then serve it with your preferred backend:

- **LM Studio** — easy GUI path; MTP may be enabled by default on supported models.
- **Ollama** — simple CLI path.
- **vLLM** — maximum configurability on Linux; use when you want OpenAI-compatible HTTP serving and custom flags.

Example LM Studio download + serve (24GB+ recommendation):

```bash
lms get qwen/qwen3.6-27b
lms load qwen/qwen3.6-27b --context-length 65536
lms server start
```

Example Ollama pull + context setting:

```bash
ollama pull qwen3.6:27b
ollama run qwen3.6:27b
# then: /set parameter num_ctx 65536
```

Use the **exact** model handle or tag as the OpenClaw Model ID.