Skip to content

About

llminfo IQ

llminfo IQ combines benchmark results to compare models of the same type. A score of 100 is average for that group.

Each 15 points is one standard deviation. Current frontier models score around 150–175. Chat models are compared with chat models, and embedding models with embedding models. The scores draw on ~90,000 benchmark results, and each one shows how much data sits behind it.

Each benchmark is weighted by how much signal it actually carries. The method is open. We publish the full per-benchmark weight table with every refresh. Every refresh, we also check the index against independent aggregate ratings. The main check is the Epoch Capabilities Index (Spearman ρ ≈ 0.98). The second is the Artificial Analysis Intelligence Index, as relayed by third-party sources. Any quoted agreement with an Artificial Analysis figure names the source that relayed it, never Artificial Analysis as our own source.

Last calculated 22 Nov 2026. A model's score can change over time as new benchmark results arrive and the pool of models shifts. It's always shown with its date.

Current frontier: about 150–175557085100115130145160175100 average for its type
The ruler shows 55 to 175. That is the range drawn, not a limit of the score. Each tick is 15 points, one standard deviation.

What it is not

  • Not a measurement of any single benchmark. It combines results from many tests. A high-IQ model can still lose a particular benchmark to a lower-IQ one.
  • Not comparable across model types. Chat models are ranked among chat models, embedding models among embedding models. An embedding IQ of 130 and a chat IQ of 130 are unrelated claims.
  • Not a quality guarantee for a domain. A single number can't capture that a model is unusually good at coding for its general ability. Coding has its own index, Coding IQ, fitted from coding benchmarks on a separate scale. The two numbers aren't comparable, and their difference doesn't measure specialisation.
  • Not immune to gaps in the data. A score with little data behind it carries real uncertainty. Models tested only on niche community boards are pulled toward the average on purpose.
  • Not human IQ. The name borrows the scale convention (100 ± 15) for readability, nothing else.

How much data sits behind a score

Select a score in the table to see how much data sits behind it. A score with little data behind it is a rough estimate and is shown as one.

  • highPlenty of data behind it: 10 or more benchmarks, including at least 2 from trusted-tier sources.
  • mediumA fair amount of data behind it: 5 or more benchmarks.
  • lowLittle data behind it: 3 or 4 benchmarks, shown as a rough estimate (≈ N, no decimals).

Current weight table

The benchmarks that drive the chat score most right now, by measured weight. A benchmark’s weight is how well it separates strong models from weak ones. Weights are re-measured from the data on every refresh. Nothing is hand-picked, so this table moves.

BenchmarkWeightModels
Vals SRE Bench4.3434
Vals Terminal-Bench-Science4.1846
Blueprint Bench 2 Raw score4.0340
SWE-bench Verified (Epoch)3.6131
Vals SWE-bench Multimodal3.6021
SWE-Bench Pro3.4892
Blueprint Bench 2 Score3.4640
OpenHands Index: swe-bench3.4631
Vals ProgramBench3.4264
NYT Connections – Older models (Lech Mazur)3.3777
VideoMMMU3.3327
Vals SWE-bench3.3383
LiveCodeBench Easy3.3216
Deep20Bench success rate3.2728
Multi-SWE-bench JavaScript Verified3.2711
MathVista-Mini3.2423
Vals EMB3.1473
Multi-SWE-Bench3.1210
Vals AIME3.0981
GDPval3.0910
Arena Agent Tool Hallucination3.0862
Vals Router Index3.0824
Humanity's Last Exam (no tools, text-only)3.0415
FrontierMath Tier 43.0147
GSO2.9930
METR Time Horizons2.9935
SWE-bench Multilingual2.9758
FrontierCode 1.12.9741
GDM-MRCRv2: 8 needle 8k2.9668
Vals Index2.9340
BrowserGym WorkArena-L22.9211
Vals EMB v1.12.8944
Vals LiveCodeBench2.89127
LiveCodeBench Medium2.8916
INCLUDE-base-44 0-shot greek (Speakleash)2.8915
BFCL-v32.8816
LiveBench theory_of_mind2.8551
LiveBench zebra_puzzle2.8451
VideoMME w sub.2.8410
PutnamBench lean-wsolution solved2.8312
Other model types (embedding, image generation, video generation, text to speech)

embedding

BenchmarkWeightModels
MTEB(fas, v2) - Clustering1.8812
ViDoRe(v3) - Public split1.826
ViDoRe(v3)1.806
MTEB(fra, v1)1.7217
MTEB(fra, v1) - Clustering1.6522
MTEB(fra, v1) - Mean (TaskType)1.6517
BEIR-NL1.635
MTEB(kor, v1)1.6321

image generation

BenchmarkWeightModels
UniGenBench++ (Chinese Long) - Color2.5125
UniGenBench++ (Chinese) - Animal2.4428
UniGenBench++ (Chinese Long) - Layout-Overall2.4025
UniGenBench++ (Chinese) - Overall2.3728
UniGenBench++ (Chinese Long) - Material2.3325
UniGenBench++ (Chinese Long) - Attribute-Overall2.3125
UniGenBench++ (Chinese) - Feature matching2.2828
UniGenBench++ (Chinese) - 2D2.2728

video generation

BenchmarkWeightModels
HumanScore2.189
HumanScore Kinematic2.109
Design Arena Elo (models/multitovideo)1.8914
Physics-IQ Verified1.7310
Arena ELO Text to Video1.3839
llm-stats Video Generation Arena1.3824
Arena ELO Video Edit1.3610
Design Arena Elo (models/imagetovideo)1.3326

text to speech

BenchmarkWeightModels
Design Arena Elo (models/audiorealism)1.3210
Pashto ASR FLEURS ps_af CER (zero-shot)1.247
Vals VoiceCodeBench1.2310
MAEB(v1, audio-only) - Mean (TaskType)1.235
EmergentTTS-Eval Syntactic Complexity WER1.167
EmergentTTS-Eval Questions WER1.157
Pashto ASR FLEURS ps_af WER (zero-shot)1.117
Pashto ASR CV24_filtered WER (zero-shot)1.097