STOP GUESSING. START RANKING. The Independent LLM Standard.

Built for developers, not marketing teams. Every score is reproducible, open-source, and conflict-of-interest-free.

1. The Five Dimensions

Benchmarks (70%)

Evaluates logical reasoning, coding, math, and knowledge.

Sources: HuggingFace Evals, Open LLM Leaderboard V2. A 90/100 means top-tier reasoning. Limitation: Does not capture creative writing preference.

Efficiency (5%)

Throughput, VRAM usage, and parameter-to-performance ratio.

Sources: Context length metadata, param count. A 90/100 means runs fast on consumer GPUs. Limitation: Static estimates, not real-time profiling.

Community (10%)

Usage, momentum, and mindshare.

Sources: HF Downloads, likes. A 90/100 means mass adoption. Limitation: Can be skewed by early hype or bots.

Recency (15%)

Time since release and update frequency (180-day half-life).

Sources: Last modified dates. A 90/100 means updated recently. Limitation: Penalizes stable, completed models over time.

Reproducibility (0%)

Open weights, clear license, verified origin. Reserved for future weighting.

Sources: Hub metadata, safetensors presence. A 90/100 means fully open (MIT/Apache) and safe.

2. Benchmark Coverage Table

Benchmark	Domain	Source	Weight	Notes
MMLU-Pro	General knowledge	HuggingFace Evals	20%	...
GPQA Diamond	PhD-level reasoning	idavidrein/gpqa	20%	...
HLE	Expert-level	...	15%	Humanity's Last Exam
GSM8K	Math word problems	...	10%
HumanEval	Code generation	openai/...	10%
BBH	Big-Bench Hard	...	8%
IFEval	Instruction following	...	7%
MuSR	Multi-step reasoning	...	5%
MATH	Advanced math	...	5%
ARC-Challenge	Science reasoning	...	fallback
HellaSwag	Commonsense NLI	...	fallback
TruthfulQA	Factual accuracy	...	fallback
WinoGrande	Winograd schema	...	fallback

3. Normalization & Confidence

Raw benchmark values from HF are 0.0-1.0, we multiply by 100. Frontier benchmarks (MMLU-Pro, GPQA etc.) take priority. When only classic benchmarks found: 0.85x confidence penalty, capped at 75/100. Coverage confidence: high/medium/low based on how many benchmarks found.

4. Tier System

Tier	Score Range	Current Examples
S	90-100	(none yet — GPT-4 class)
A	80-89	gemma-4-31B-it (82.97), Qwen3.5-9B (81.52)
B	70-79	DeepSeek-R1 (78.3), phi-4 (72.89)
C	60-69	gpt-oss-20b (69.81)
D	<60	Legacy models

5. ELO Comparison Formula

P(A beats B) = 1 / (1 + 10^((ELO_B - ELO_A) / 400))

Example: Qwen3.5-9B (81.52) vs DeepSeek-R1 (78.3)
• ELO_A = 800 + 81.52*8 = 1452, ELO_B = 800 + 78.3*8 = 1426
• P(Qwen beats DeepSeek) = 1 / (1 + 10^((1426 - 1452) / 400)) = 0.537 = 53.7% win probability

6. Extended Metadata (10 signals)

context_window, vram_tier, license_score, finetune_friendly, multilingual, safety_score, update_velocity, inference_coverage, community_momentum, hub_completeness.

7. What We Don't Measure (Honest Limitations)

Human preference (requires live inference infrastructure)
API latency and cost per token
Alignment and safety (beyond TruthfulQA)
Benchmark contamination (we can't verify if models saw test data)
Dialect/regional language performance

8. Changelog

Version	Date	Changes
2.0.0	2026-08-13	10 extended metadata signals, Shields.io endpoint, pricing page
1.1.0	2026-08-13	GitHub Pages CDN, 71 models, HuggingFace Space
1.0.0	2026-08-13	Initial: 5D scoring, ELO, SVG badges, 27 tests

9. Cite ModelRank

@software{modelrank2026,
  author = {ModelRank Team},
  title = {ModelRank: Composite Scoring and Embeddable Badges for Open-Weight AI Models},
  year = {2026},
  url = {https://github.com/rankmodel/rankmodel.github.io},
  license = {MIT}
}

ModelRank Scoring Methodology