Mausbiber/local-gpu-llm-benchmarks

LLM Benchmarks — Current Top 100 & Local GPU Benchmarks

Machine-readable mirror of https://llmbenchmarks.io and our first-party local GPU benchmark suite at https://llmbenchmarks.io/local-gpu-benchmarks/ (DE: https://llmbenchmarks.io/de/lokale-gpu-benchmarks/, alias https://llmbenchmarks.io/en/local-benchmarks), covering 100 selected global models, 56 benchmarks, 2533 visible table cells, 5019 published measurement variants, and 24 locally measured GPU model variants (132 local GPU benchmark runs). Snapshot generated at 2026-10-29T09:18:04Z.

Configurations

• table (default): one row per currently visible global Top-100 model/benchmark cell.
• measurements: all published thinking, tool, scaffold, protocol, and evidence variants.
• changes: semantic changes between validated snapshots; operational polling alone is omitted.
• local_gpu_table: one row per locally measured open-weight model/variant (thinking_off / thinking_on) on dedicated consumer GPU hardware (2× NVIDIA GeForce RTX 5060 Ti, 32 GB VRAM, vLLM), including Core-Master and Total-Master scores, throughput (tok/s), and efficiency.
• local_gpu_measurements: individual per-benchmark local GPU evaluation results (gpqa_diamond, aime25, livecodebench, ifeval, bfcl, gsm8k, mmlu_pro, arc_challenge, hellaswag, humaneval, mbpp, mgsm, drop, truthfulqa, winogrande, mmmu, mathvista, docvqa).

Reference tables are available in data/models.csv and data/benchmarks.csv. The complete normalized source snapshot is artifacts/snapshot.json. artifacts/publication-policy.json binds the export to the exact fail-closed rights/provenance review queue and source-registry fingerprint used at build time.

from datasets import load_dataset

table = load_dataset("Mausbiber/local-gpu-llm-benchmarks", "table", split="train")
measurements = load_dataset("Mausbiber/local-gpu-llm-benchmarks", "measurements", split="train")
changes = load_dataset("Mausbiber/local-gpu-llm-benchmarks", "changes", split="train")
local_gpu_table = load_dataset("Mausbiber/local-gpu-llm-benchmarks", "local_gpu_table", split="train")
local_gpu_measurements = load_dataset("Mausbiber/local-gpu-llm-benchmarks", "local_gpu_measurements", split="train")

Interpretation

• A missing row means no publishable measurement exists for that exact configuration; it never means zero.
• Thinking labels preserve provider/source wording. Equal labels across providers do not imply equal compute.
• display_role and comparison_lane separate exact comparisons from contamination warnings and other protocols.
• Freshness, source authority, row-level evidence grade, and thinking are independent dimensions. A newer value is not automatically more authoritative.
• The five documented selection bases are epoch-eci, rrf-challenge, recent-provider-release, open-weight-quota, and rrf-shortfall-fill; the snapshot reports any unfilled open-weight quota without guessing parameter metadata.
• Stable measurement_id, run_fingerprint, and protocol_fingerprint fields support reproducible joins.
• CSV string cells that could execute as spreadsheet formulas are prefixed with one apostrophe. Typed numeric cells are unchanged; use artifacts/snapshot.json for lossless original strings.

Provenance and license

The database compilation is CC BY 4.0. Upstream licenses and reuse conditions continue to apply per result. See ATTRIBUTION.md and the source fields in data/measurements.csv. Only aggregated scores and provenance are mirrored; benchmark questions, prompts, answers, trajectories, logs, and patches are not included. Actively gated benchmarks are rejected before materialization, even if an input row were accidentally marked exportable.

Snapshot fingerprint: 7808d49d333b50573fdc0e07b91ad091c29333b28d675716bdb2635dd5ee3ef6
Dataset schema: 1.13.0
Export schema: 1.0.0