tekosML/qwen36-cross-platform-benchmark

Qwen3.6-35B-A3B cross-platform benchmark dataset

Complete sanitized evidence for the Qwen3.6 GX10 versus M2 Max benchmark.

Headline results

• At 128K, median cold-prompt TTFT was 45.70 s on GX10 and 552.79 s on M2 Max, a 12.10x difference.
• Near 256K, GX10 completed and strictly passed 12/12 requests. M2 Max completed 9/12 and strictly passed 7/9 completed requests.
• On the same Q4_K_M coding control, MTP improved median decode 26.87% on GX10 CUDA and regressed 19.35% on M2 Max Metal.

Read the final technical report for configuration details, quality denominators, runtime failures, and limitations.

Contents

• processed-results/results.parquet: 429 normalized result rows for analysis.
• processed-results/results.jsonl: the same generated table as JSONL.
• raw-results/: append-only outputs, resolved configs, and environment records.
• configs/: staged benchmark configurations.
• tasks/: deterministic coding, tool, retrieval, and Puzzle contracts.
• reports/: final reports and the agent working log.
• reproductions/: retained vLLM and llama.cpp failure evidence.
• figures/: charts generated from the raw dataset.

Privacy transformation

The exact private archive is retained offline. This public dataset replaces local macOS/Linux usernames, personal home paths, hostnames, and Tailscale IP addresses with stable pseudonymous placeholders. Model outputs, timings, task inputs, quality results, model revisions, runtime revisions, and launch flags are unchanged.

Run IDs and sample IDs are preserved so published reports and row references remain stable. Because endpoint and filesystem fields were redacted after collection, a public config.resolved.json will not reproduce the original config hash byte-for-byte. Those fields are transport metadata, not benchmark variables.

Reproducibility

git clone https://github.com/sxuff/qwen36-cross-platform
cd qwen36-cross-platform
python -m pip install -e '.[dataset]'
python scripts/fetch_dataset.py --output data/

No model weights are included. Use the immutable upstream model and checkpoint revisions recorded in the reports.

License

Original benchmark code and materials are Apache-2.0. Model checkpoints and third-party runtimes retain their upstream licenses.