Qwen3-Coder 30B on a Surface Laptop Ultra: run it on RTX Spark with Ollama

On the Microsoft Surface Laptop Ultra, the Qwen3-coder:30B model demonstrated impressive performance when utilizing the NVIDIA RTX Spark N1X GPU with 64 GB of LPDDR5X-9400 memory. The model achieved a token decoding speed of 90.9 tokens per second, nearly twice the speed of the Qwen3.8-27B model on the same chip. Both laptops shared a 64 GB LPDDR5X-9400 memory setup, with the GPU receiving its dedicated 31,232 MiB.

When running the Qwen3-coder model locally with Ollama on Windows, without Docker or vLLM, the Surface Laptop Ultra exhibited exceptional performance. The long prompt challenge saw a significant slowdown, with the model taking 146 seconds to process the first token and subsequently decoding at 19.7 tokens per second. The GPU's average power consumption was 33 W, peaking at 75 °C.

The software setup was straightforward, involving a single signed installer and a single Ollama pull, with no need for WSL or Docker environments. The Qwen3-coder model, consisting of 30.5B parameters with 8 of 128 experts active per token, required a 102,400-token context that comfortably fit within the GPU's memory capacity of 26.7 GiB.

The decoding speed dropped dramatically between short code prompts and longer prompts, falling about 4.6 times between a 20-token prompt and a 100K-token prompt due to the attention cache growing with every token. Ollama silently truncated prompts longer than its context window to approximately half of the original length, without providing any indication.

The Qwen3-coder model is ideal for developers seeking a local coding model for chat, edits, and short agent turns, without incurring per-token costs or managing container stacks. The Surface Laptop Ultra provided a powerful and efficient platform for this specific use case, offering an ultra-fast experience for short, interactive coding requests.