llmrun/llmrun-composite-score

llmrun Composite Score

A single capability score for 243 language models, 100 open-weight and 143 proprietary, fitted across 45 public benchmarks. Each score includes a 90% range, while most models also receive up to five skill scores: reasoning, coding, math, science, and agents.

This is the data powering the llmrun Composite Score leaderboard. You can find the method, every change made, and the full benchmark list on the methodology page.

Browse the rankings for either all models or just the open ones in the llmrun Composite Score Space.

What's inside

Config	Rows	What a row is
scores (default)	243	One model in the latest snapshot, open-weight or proprietary
scores_by_setting	341	One proprietary model at one reasoning-effort setting
boards	45	One benchmark in the latest fit, with its difficulty and discrimination
history	grows	Every model in every snapshot, with snapshot_date
scores_by_setting_history	grows	Every proprietary setting in every snapshot
boards_history	grows	Every benchmark in every snapshot

from datasets import load_dataset

scores = load_dataset("llmrun/llmrun-composite-score", "scores", split="train").to_pandas()
open_coders = scores[scores.open_weights].dropna(subset=["coding_score"])
print(open_coders.sort_values("coding_score", ascending=False)[["model", "coding_score", "coding_low_90", "coding_high_90", "params_total_b"]].head(10))

How the score works

We run no models ourselves; we simply use results already published by independent leaderboards: the Epoch AI Benchmarking Hub, LiveBench, the SWE-bench leaderboards and MMLU-Pro.

Each result is rescaled from the benchmark's random-guess floor to its maximum. A two-parameter item response theory model then estimates one hidden ability per model along with a difficulty and discrimination for each benchmark. Because every benchmark has its own difficulty level, a model tested only on easy ones won't climb above those assessed on harder tasks, which is where plain averaging fails.

Ability appears as score = 100 + 25 × θ. A score of 100 represents the average model in the 2026-Q4 reference group; 25 points equals one standard deviation, and there's no ceiling. The 90% range shows the score at θ ± 1.645 standard errors. When fewer or less informative benchmarks are used, that range widens, so two models with overlapping ranges should be treated as roughly equal in ability.

A benchmark counts only while it has results from the last 365 days; older saturated benchmarks drop out automatically. Relative ratings like LMArena Elo aren't included. A model receives an overall score once it has results from at least 4 benchmarks spanning at least 2 categories, including reasoning or coding, or from 3 benchmarks that narrow the range to 15.6 points or fewer. In either case, data must come from at least 2 different benchmarks. Models under that threshold aren't included in the tables.

Proprietary models and effort settings

Sources often list a proprietary model multiple times, once for each reasoning-effort setting, such as GPT 6 Astra at low, high, xhigh, and max levels. Every setting is fitted independently, with its per-setting scores recorded in the scores_by_setting table.

In the main table, every proprietary model occupies its own row. For each benchmark, we record the result from the highest reported setting; if no specific setting is stated, we treat that as the provider default, which falls between medium and high. The combined row scores against benchmarks already calibrated using the per-setting fit, so it doesn't shift any open models. The setting_names column lists the settings folded into each row.

Snapshots and versions

New models and new results roll out weekly, making the score a moving target. Each snapshot captures a full copy of the table as it stood when fitted, with older versions remaining untouched. data/latest/ always points to the latest entry, while data/history/ stores them all. Every snapshot carries the tag v<date> inside the repository, letting you pin an exact version via revision=.

The scale stays locked to the 2026-Q4 reference point. Benchmarks added afterward are calibrated against it, and earlier benchmarks keep their original parameters. Inside any single methodology_version, a model's score shifts only when its own results change; this means you can compare scores across different snapshots directly. If the method changes, the version string updates too, with details listed in the changelog on the methodology page. Always compare models within one version.

Columns: scores

Column	Meaning
snapshot_date	Date of the benchmark data behind the fit
methodology_version, scale_reference	Method version and frozen scale reference
rank, rank_open	Rank among all models, and among open-weight models only
model, open_weights	Model name; true for open-weight models
publisher, family	Publisher and family (open-weight models)
score, score_low_90, score_high_90	Overall score and its 90% range
benchmark_count, categories_present	How many benchmarks and categories the score rests on
<skill>_score, <skill>_low_90, <skill>_high_90	Skill scores for reasoning, coding, math, science and agents; empty when the model has too few results in that skill
settings, setting_names	Proprietary models: how many effort settings were folded into the row, and which
params_total_b, params_active_b	Total and active parameters in billions; active is filled for mixture-of-experts models
context_length, license, release_date, hf_repo	Open-weight models, from the Hugging Face repository
llmrun_url	Open-weight models: model page on llmrun

scores_by_setting has the same score and skill columns, with model (the row in scores) and setting_name (the setting as the source lists it).

Columns: boards

Column	Meaning
board, name, category	Benchmark id, name and category
source, source_url	Where the results come from
irt_discrimination_a, irt_difficulty_b	Fitted IRT parameters
observations	Number of results in the fit
scale_anchor	true for benchmarks in the frozen reference, false for benchmarks calibrated onto it later
llmrun_url	Benchmark page on llmrun

Limitations

The benchmarks are mostly in English, with no vision or multilingual skill yet. We inherit whatever the upstream leaderboards publish, including their harness choices and any contamination issues. Because testing on only a few benchmarks yields a wide performance range, always check that range before looking at the rank.

Sources and attribution

Scores come from results published by Epoch AI Benchmarking Hub, LiveBench, the SWE-bench leaderboards and MMLU-Pro. We don't redistribute raw per-benchmark numbers here; check each source for those figures and their terms. When using this data, please credit the original benchmarks.

Citation

@misc{llmrun_composite_score,
  title  = {llmrun Composite Score},
  author = {{llmrun}},
  year   = {2026},
  url    = {https://llmrun.dev/benchmark/llmrun-score},
  note   = {Dataset: https://huggingface.co/datasets/llmrun/llmrun-composite-score}
}

More

• Interactive chart on Hugging Face Spaces
• Leaderboard and methodology
• Write-ups: why averaging benchmarks gives the wrong leaderboard, fitting one score across sparse benchmarks with IRT