Qwen36 Mtp Turbo Kv Analysis: Runtime Benchmarks for Quantized Models

A curated analysis artifact comparing inference runtimes for the Qwen3.6-35B-A3B model. It contains results from experiments on Windows CUDA, comparing clean MTP llama.cpp, QuinsZouls llama-next TurboQuant, and Atomic TurboQuant runs under a fixed 64k context. The dataset was created by sjakek and last updated on 2026-05-15.

Use Cases

• Compare inference speed across different quantization methods based on the described TurboQuant variants.
• Benchmark local LLM inference performance on Windows CUDA hardware based on the experimental setup.
• Analyze the impact of context length and CPU offload settings on runtime based on the fixed 64k context and MoE CPU offload parameters.

Strengths

• Benchmark runs were conducted under a fixed 64k context, MoE CPU offload, and Unsloth-aligned sampling settings, enabling direct comparison.
• The repository is curated to include only completed, comparable runs, filtering out incomplete or incompatible data.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The dataset description is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Local inference experiments on Windows CUDA.

Freshness

Last updated 2026-05-15 15:31:11; freshness should be verified.