MTEB English v2 — Massive Text Embedding Benchmark
MTEB(eng, v2): text embedding quality across 41 tasks in the captured official suite.
Official Mean (Task) × 100, on a 0–100 scale; higher is better. Only complete canonical runs are ranked. Mean (Task Type) and other suites are separate. Task count is not question count. Suite hash, dimensions and training overlap are preserved per run; unpublished model revisions and evaluation dates stay unknown.
Current recorded best scores, ordered by model release date.
Hover or tap to inspect. Keyboard: ← / → to move, Home / End to jump, Esc to reset.
Qwen3-Embedding-8B: 75.2, released Jun 5, 2025
Tested by MTEB official leaderboard · All results are without tools.
Harness, budget and task set can change the same model’s score.
Qwen3-Embedding-8B leads with 75.2, measured independently (MTEB official MTEB(eng, v2) leaderboard).
Independent: the benchmark’s own leaderboard or a third-party evaluator. Provider-reported: the model maker’s model card, system card or launch post. When sources report the same result, the board shows the official leaderboard first, then the provider, then third-party tests.
Compare the same agent harness, effort setting and test budget. Small score differences may not be meaningful. Use cost per task from the scored run; token prices alone omit retries and agent overhead.
The dataset license covers the test itself; each score source sets its own terms.
Published by MTEB community / embeddings-benchmark · No license recorded; check the publisher’s terms
Unless a license is listed, each source’s own terms apply.
Published by MTEB community / embeddings-benchmark · CC0-1.0 · Score-data terms
Published by Google DeepMind