Source: Minish (Thomas van Dongen and Stephan Tulkens, the Model2Vec authors), release post 2026-11-18 and the model card on Hugging Face. MIT licence. All scores below are the authors' own, run with mteb 2.27.1.
What it is
blink-multilingual-64M is a static embedding model. There is no transformer at inference time: the whole model is one 250,000 × 256 table of token vectors (64M parameters, 256 MB in float32). To embed a text you tokenize it, look up each token's vector and take the mean. That makes encoding a memory lookup plus an average, orders of magnitude cheaper than running a transformer encoder, and it runs anywhere (Python via model2vec or Sentence Transformers, Rust via model2vec-rs).
It replaces the team's earlier potion-multilingual-128M at half the size, and follows an English-only blink-8M released a week earlier.
Headline numbers
MTEB Multilingual v2, all 131 tasks, every model compared at 256 dimensions:
| Model | Mean (task) | Retrieval | STS | Bitext |
|---|---|---|---|---|
| blink-multilingual-64M | 51.60 | 47.52 | 67.47 | 57.03 |
| bekko-static-embedding-v1 (truncated to 256d) | 48.37 | 47.62 | 63.72 | 37.30 |
| potion-multilingual-128M | 47.67 | 41.22 | 61.15 | 40.72 |
| static-similarity-mrl-multilingual-v1 (256d) | 46.88 | 39.49 | 63.67 | 48.50 |
It also beats bekko at bekko's full 1024 dimensions on the overall mean (50.18), though not on retrieval (49.96).
On the 96 non-English tasks it leads by 4 points (51.38 vs 47.39). On the 35 English-only tasks it is tied with bekko.
On small retrieval suites (HAKARI Nano), it leads multilingual NanoBEIR (41.18 vs 39.62), ties on English, and trails bekko on NanoMIRACL (65.41 vs 66.52).
The honest reading: a big jump over its predecessor and a clear lead on bitext and STS, but retrieval is a tie with bekko, not a win.
How it was built (six steps)
Tokenizer. A 250k uncased WordPiece vocabulary that contains every letter of 57 scripts and splits long character runs, so nothing becomes
[UNK]. A stock WordPiece lost 10% of Thai and 23% of Lao characters to[UNK]. 20k Chinese and 10k Japanese words are added as whole tokens.Initialization. Each token is embedded by bge-m3, reduced to 256 dimensions with PCA, and weighted by token frequency (SIF).
Distillation (Tokenlearn). The table learns to reproduce bge-m3's sentence embeddings for 4M multilingual C4 texts, with a cosine loss.
Contrastive training. Following the bekko recipe: MNRL at temperature 0.05 on 64M pairs (mMARCO, MIRACL, NLI, bitext and more), one data source per batch, hard negatives where available, trained natively at 256 dimensions.
Weight averaging. Final weights are 0.8 × contrastive + 0.2 × distilled.
Language codes. Contrastive training on translations erases which language a text is in. Each language gets a fixed random vector, and each token is nudged towards the languages it occurs in. Language-ID tasks jump (IndicLangClassification 26.3 → 64.9) for a 0.07-point average cost elsewhere. The model is still a plain table.
Architecture: where a static model fits
query / doc chunk -> tokenize (250k WordPiece) -> look up 256d rows -> mean pool + L2 norm
docs -> vector index (256d)
query -> cosine top-k over index -> optional transformer re-ranker -> results
A static model is a candidate generator and a cheap similarity function, not a full replacement for a transformer encoder. The common shape is a two-stage pipeline: the table embeds the whole corpus and every query cheaply, then a transformer re-ranker or a stronger embedder scores the top few hundred hits.
Tradeoffs
Speed and footprint vs quality. You give up context sensitivity: a bag of token vectors cannot tell "dog bites man" from "man bites dog", and negation is weak. In exchange, encoding is effectively free, runs on CPU or at the edge, and re-embedding a large corpus is cheap.
256 dimensions. Small vectors cut index memory to a quarter of a 1024d model's. The comparisons truncate the rivals to 256d, which favors a model trained natively at 256.
Distilled from bge-m3. Quality is bounded by the teacher's view of the world, and the vocabulary is fixed at training time. New jargon becomes subword soup.
Mean pooling. Long texts average out to a blur. Chunk before you embed.
Failure modes to watch
Word order and negation. Queries whose meaning hinges on order or "not" will retrieve the wrong side. A re-ranker is the usual guard.
Benchmark mix. MMTEB is weighted towards English (27% of tasks), which the authors call out and correct for. Your own corpus still needs its own eval.
Training overlap. Both blink and bekko train on MIRACL, so MIRACL scores are partly in-distribution.
Mixed-language corpora. The language-code shift helps language ID, but if you want cross-lingual retrieval, check that it doesn't push same-meaning, different-language texts apart for your pairs.
Review checklist
Is the static model a first stage with a re-ranker behind it, or the only scorer? If only, test order- and negation-sensitive queries.
Was it evaluated on your own queries and corpus, not just MMTEB?
Are texts chunked to a size where mean pooling still means something?
Is the 256d index benchmarked against your current embedder's recall at the same k?
For cross-lingual search, are translation pairs from your domain checked?
Is the model version pinned, given two releases in one week?
Related
R106 txtai: vectors, SQL and a graph in one local index (a local store that can host a cheap embedder).
R127 GraphRAG Zero: graph-guided retrieval with no LLM calls at index time (another push to make indexing cheap).
R010 PageIndex: reasoning-based retrieval for long documents (the opposite bet: no embeddings at all).
R054 eleven coding harnesses that don't use embeddings to find code (where embeddings are not the right tool).