RTX 5090 Benchmarks: Speed and Quality for Quantized LLMs

Benchmarks for quantized large language models on the NVIDIA RTX 5090 32GB GPU, measured with llm-bench-rig. The dataset includes generative quality evaluations via llama-server chat completions, using a custom methodology that separates reasoning and non-reasoning model results. It was created by witcheer and last updated on July 8, 2026.

Use Cases

• Compare inference speed of quantized models based on benchmark results for the RTX 5090.
• Evaluate generative quality of reasoning versus non-reasoning LLMs based on the described split methodology.
• Select optimal model quantization levels for deployment on specific GPU hardware based on performance trade-offs.

Strengths

• Benchmarks are performed on a specific hardware configuration (NVIDIA RTX 5090 32GB).
• Quality evaluation methodology is described, including a custom approach that separates reasoning modes.

Limitations

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

Provenance

Source

huggingface

Collection Method

Benchmarks measured with llm-bench-rig.

Freshness

Last updated 2026-07-08 13:25:36; freshness should be verified.