Skip to content
Local AI Wizard
Your machine
Theme

How do I read LLM benchmark scores?

LiveBench, LMArena and MTEB measure different things on different scales, so the site never mixes them.

Getting started · updated

Short answer

Compare a score only with others from the same source and release. Each source has its own scale, so the site never averages them.

  • LiveBench: Test questions, scored 0–100.
  • LMArena: A rating from blind votes by people comparing answers.
  • MTEB: Scores embedding models only.

Counted from the site's bundled LiveBench, LMArena and MTEB data. Methodology

The three sources

What each one measures, its scale and how many reviewed models carry it today.

Quality sources shown on the site, by what they measure and their coverage.
SourceWhat it measuresScaleRelease usedReviewed models with a value
LiveBenchRight answers on new questions in math, coding, reasoning, data analysis, language and instruction following0–100, higher is better2026-06-2522 / 60
LMArenaWhich of two anonymous answers people prefer, turned into a ratingRating with a ± interval; reviewed models span 1,111 to 1,4802026-10-2225 / 60
MTEB (Multilingual, v2)How well an embedding model finds and groups related text, averaged over its tasks0–100, mean per taskResults dataset16 / 20

Sources: (LiveBench 2026-06-25, read Sep 26, 2026) (LMArena 2026-10-22 · CC BY 4.0, read Oct 25, 2026) (MTEB (Multilingual, v2) · Qwen3-Embedding 8B, read Oct 25, 2026)

One model, two scores

LiveBench 2026-06-25
75.3 / 100Qwen3.8 27B scored 75.3 out of 100, the average of its category results.
LMArena 2026-10-22
1,437 ± 5Position 102 of the release, from 26,238 votes; the ± is the interval the leaderboard publishes.

7 reviewed text models carry both. The two numbers are about different things (test answers versus people's preference) and on different scales, so a “75” and a “1,437” cannot be added, averaged or converted into each other. The site gives each its own chart lens and its own table column. Among reviewed models, the highest LiveBench score belongs to DeepSeek V4.1 Flash · max; see Qwen3.8 27B for both scores on one page.

Why they are never mixed

Different questions
Accuracy is not preferenceA model can win votes with a pleasant style and still miss test answers, or the other way round.
Different scales
0–100 vs an open-ended ratingAn Arena rating has no maximum and moves when new models join; a percentage does not.
Different releases
2026-06-25 · 2026-10-22Scores are compared only within the same release; the release date travels with every number.
Hosted, not your file
Tested on the full modelBoth text sources test the publisher's model, not a Q4 file on your machine, so neither says how a quantized file scores.

When there is no score

A dash means the model has no result in that release, and the dash carries the reason. It is never a zero and never a score borrowed from a bigger or smaller sibling. 20 reviewed text models have neither text score; their pages still show memory, files and where they run.

For search and retrieval, embedding models are scored by MTEB instead; the top reviewed result is Qwen3-Embedding 8B (70.6 / 100).