OpenClaw + Qwen3.6, Running Locally on DGX Spark
/ 3 min read
I set up a local agent on my DGX Spark: OpenClaw on top, Qwen3.6-35B-A3B served by vLLM underneath. Short version: 92.6 to 93.0 tok/s single stream with MTP speculative decoding, 62.3 without it. Speculation buys about 1.5x. The reason it works at all is the same one from my Gemma 4 benchmarks. Decode is a memory bandwidth problem, and a model that activates roughly 3B parameters per token does not move much memory.
The workload I have in mind is polling: check a page or a file every few minutes and do a small thing when something changes, all day. On a cloud model that burns paid tokens forever. Locally it burns electricity I am already paying for. It also decides what the agent can touch. A 35B local model in an open harness does not get my email or my accounts. It works with files I give it and the public web, and anything it drafts waits for my approval. That is the whole trust model.
- Model: Qwen3.6-35B-A3B, NVIDIA’s NVFP4 build
- Serving: vLLM v0.28.0 in Docker, FP8 KV cache, 262K context
- Agent: OpenClaw 2026.10.3, local endpoint only
- Benchmarks: single stream, 1024 token generations, reasoning tokens counted
The recipe is NVIDIA’s playbook, with two forced changes. Its GPU memory setting of 0.4 leaves too little KV cache at 262K context and the engine refuses to start; 0.5 works. Docker GPU access needs --gpus all, because the CDI flag the docs prefer does not work here.
| Config | Output tok/s |
|---|---|
| NVFP4, no speculation | 62.3 |
| NVFP4, MTP-3 | 92.6 to 93.0 |
Three warm runs, all within half a token of each other. A 512 token spot check later gave 86.7, so call real usage 85 to 95.
The number floating around for this config is roughly 105 tok/s. I got 93. That is about ten percent under, and under the 100 bar I had set for myself. It is also three times what I called unusable in the Gemma post, and for short mechanical polling generations I doubt the gap matters.
What I have not measured is quality. I haven’t run SWE-bench on this setup, or any task suite of my own. The vendor numbers for this model sit in the low 70s on SWE-bench Verified, and vendor numbers are vendor numbers. Until I have my own data, it only gets work I can verify afterwards.
The agent itself passed a basic smoke test: it summarized a file from its workspace, knew it was running on the local model, and still worked after a vLLM restart.
For now, my advice is the NVFP4 weights with MTP on, and expect about 90 rather than 105. The speed question is settled. The trust question is not.