Ivan Fioravanti ran a set of llama.cpp benchmarks for Qwen3.6-35B-A3B-UD-Q8_K_XL on DGX Spark. The results are straightforward:

- 2k context: 1573 prompt tokens/s, 79 generation tokens/s

- 8k context: 79 generation tokens/s

- 64k context: 58 generation tokens/s

- 256k context: 31 generation tokens/s



Speed drops from 79 tokens/s to 31 tokens/s, which isn't a steep decline, but TTFT (Time to First Token) spikes to 97 seconds at 256k context, and TPOT also rises to 29 seconds. The cost of long-context inference is still clearly visible.

MiaAI Lab prepared a one-click startup script for this model, specifically designed for DGX Spark and machines with 96-128GB of VRAM. It handles automatic GGUF file download, health checks, background running, and OpenAI-compatible API, all with just one `./start.sh` command. The GitHub repository also thoughtfully includes stop.sh and detailed environment variable configuration.

> One netizen asked why vLLM wasn't used? Mia replied: GGUF runs better on llama.cpp, and this 35B GGUF version outperforms other formats.

An even more interesting finding comes from another coding test. Topi Santakivi conducted a comparison of 11 models, using two non-trivial coding tasks (recoverable synchronization + pluggable components) to test architectural judgment and ambiguity handling capabilities. The results are:

- **Qwen 3.6-27B-8bit (local)**: 66.5/80

- **Claude Sonnet 4.5**: 65/80

- **Qwen 3.6-35B-A3B-8bit (agent-pair)**: 64/80

- **DeepSeek-V4-Flash-REAP-180B**: 62/80

The locally 8-bit quantized 27B model beat Sonnet 4.5 in overall score. The 35B-A3B agent-pair mode (one model writes code, another reviews in real time) also scored 64 points, 5 points higher than the solo 4-bit version.

The counterintuitive takeaway here is that quantization precision and inference framework may have a more critical impact on actual coding capability than parameter count. Local 27B-8bit outperforms cloud 35B-4bit, and 35B-A3B-8bit with agent-pair collaboration almost matches Sonnet's score.

Of course, the VRAM requirement is a practical constraint. DGX Spark's 128GB unified memory can just fit the 39GB weights of Q8_K_XL + 256K KV cache + mmproj + MTP draft context. A 48GB RTX 6000 would need to reduce context length or switch to a smaller quantization.

Mia included a GIF in the reply showing a blonde witch casting a spell, with the caption "You will be shocked by its performance". So far, it does look pretty impressive.

If you have a DGX Spark or a machine with 96GB+ VRAM, you can try this startup script. GitHub: [MiaAI-Lab/Qwen3.6-35B-A3B-UD-Q8_K_XL_DGX-Spark-Recipe](https://github.com/MiaAI-Lab/Qwen3.6-35B-A3B-UD-Q8_K_XL_DGX-Spark-Recipe)