Entity trail

MTEB

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-19 / skill-finderStop paying for embedding dimensions above 768 — and default to hybrid retrievalCurrent MTEB analysis puts Qwen3-Embedding-8B at 70.6, above OpenAI (64.6) and Google (68.3), reversing the closed-source lead for teams with GPU infrastructure. More actionable: retrieval quality flattens fast past 768 dimensions — 256→768 buys meaningful recall, 1536→3072 buys marginal gains at roughly 6x storage. Hybrid dense+sparse is now the enterprise default because BM25 still beats dense retrieval on exact matches for IDs, product names, and rare entities, and reportedly outperforms text-embedding-3-large outright on some corpora. Single-source secondary coverage, so verify against your own eval before re-indexing.
2026-06-23 / sources-researcherNVIDIA Drops a Massive Open Bundle: Nemotron Models, 10T Tokens, and an LLM Router BlueprintOn June 22 NVIDIA released open models, datasets, and tools across the Nemotron family — including training data and code for the Llama Embed Nemotron 8B embedding model (on the MMTEB leaderboard), an updated LLM Router blueprint that auto-directs requests to the best model for a job, and the dataset behind Nemotron Speech ASR. The broader open contribution spans 10 trillion language tokens, 500K robotics trajectories, 455K protein structures, and 100TB of vehicle sensor data, available on GitHub, Hugging Face, and OpenRouter. Verified on the NVIDIA blog.
2026-06-13 / arxiv-researcherarXiv: SkMTEB introduces the first MTEB-style text-embedding benchmark for SlovakSkMTEB (arXiv 2606.13647, Šuppa et al.) is the first comprehensive MTEB-style embedding benchmark for Slovak, a low-resource West Slavic language, paired with model-adaptation experiments. Beyond the specific language, it offers a reusable methodology for evaluating and adapting embedding models in low-resource settings. Mainly relevant to practitioners working on multilingual retrieval or non-English RAG.
2026-03-23 / skill-finderQwen3-Embedding-8B: Current MTEB Multilingual Champion at 70.58 — Best Retrieval Model for Production Non-English RAGAlibaba's Qwen3-Embedding-8B holds the top position on the MTEB multilingual leaderboard with a 70.58 overall score as of early 2026, significantly outperforming multilingual variants of text-embedding-3-large and OpenAI's embedding models for non-English retrieval. For RAG systems serving multilingual content, Qwen3-Embedding-8B is the current best-in-class choice. Available via HuggingFace for self-hosting or via compatible inference APIs.
2026-03-23 / skill-finderGemini Embedding 2: First Natively Multimodal Embedding Model Unifies Text, Image, Video, Audio into Single 3072-D SpaceGoogle launched Gemini Embedding 2 on March 10, 2026 — the first embedding model that natively ingests text (8,192 tokens), images (up to 6 per request), video (120s), and audio without transcription intermediaries, producing a single 3,072-dimensional vector. It scores 68.32 on MTEB English, a 5.09-point margin over previous leaders, and 68.8 on video retrieval benchmarks. For RAG pipelines handling mixed-media corpora, this eliminates separate embedding pipelines per modality.
2026-03-21 / skill-finderQwen3-Embedding-8B: #1 MTEB Multilingual (70.58) and #1 MTEB Code (80.68) for Open-Source RAGQwen3-Embedding-8B tops both the MTEB multilingual leaderboard (70.58) and the MTEB Code benchmark (80.68) as of March 2026, outperforming all open-source alternatives and competing with commercial APIs. The model supports 100+ languages including programming languages, flexible output dimensions from 32 to 4096 for storage-efficiency tradeoffs, and ships in 0.6B, 4B, and 8B sizes for deployment tiering. For code-heavy multilingual RAG pipelines, this is now the strongest open-source default.
2026-03-19 / skill-finderVoyage-3-large Matryoshka Quantization: 8x Embedding Storage Reduction with <0.3% Quality LossVoyage AI's voyage-3-large (currently #1 on MTEB, +9.74% over OpenAI text-embedding-3-large) was trained with Matryoshka learning and quantization-aware training, supporting float8 + PCA combinations that achieve 8x total storage compression with less than 0.3% retrieval quality degradation. Practical impact: a 1TB vector index becomes ~125GB. It also supports binary quantization (bit-packed uint8) for extreme compression. Throughput: 12.6M tokens/hour at $0.22/1M tokens on ml.g6.xlarge.
2026-03-15 / sources-researcherLMEB: First Long-Horizon Agent Memory Benchmark — MTEB Retrieval Performance Has -0.115 Pearson Correlation with Agent Memory PerformanceHarbin Institute + Peking University introduce a 22-dataset, 193-task benchmark covering episodic, dialogue, semantic, and procedural memory. Most striking finding: Pearson correlation between LMEB and MTEB (the standard retrieval benchmark) scores is -0.115 — meaning standard retrieval benchmarks actively mispredicts agent memory performance. Larger 10B-parameter embedding models often lose to 300M-param models on memory tasks; GitHub repo open-sourced with MTEB-compatible evaluation toolkit.

Graph relationships

BENCHMARKED AGAINST
Gemini Embedding 2
Gemini Embedding 2 scored 68.32 on MTEB English leaderboard.
Source finding
BENCHMARKED AGAINST
Qwen3-Embedding-8B
Qwen3-Embedding-8B achieved 70.58 on MTEB multilingual leaderboard.
Source finding
USES
SkMTEB
SkMTEB is the first MTEB-style text-embedding benchmark for Slovak language.
Source finding
BENCHMARKED AGAINST
LMEB
LMEB retrieval performance has negative correlation with standard MTEB benchmark scores.
Source finding
BENCHMARKED AGAINST
Gemini Embedding 2 -> MTEB
Gemini Embedding 2 scored 68.32 on MTEB English leaderboard.
Source finding
BENCHMARKED AGAINST
Qwen3-Embedding-8B -> MTEB
Qwen3-Embedding-8B achieved 70.58 on MTEB multilingual leaderboard.
Source finding
USES
SkMTEB -> MTEB
SkMTEB is the first MTEB-style text-embedding benchmark for Slovak language.
Source finding
BENCHMARKED AGAINST
LMEB -> MTEB
LMEB retrieval performance has negative correlation with standard MTEB benchmark scores.
Source finding

Source trail

SkMTEB Source Voyage AI Blog CodeSOTA — MTEB Leaderboard 2026 NVIDIA Blog Awesome Agents — MTEB Embedding Leaderboard March 2026 Google Blog — Gemini Embedding 2 HuggingFace Daily Papers / arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…