blink-multilingual-64M
by minishlab
A static embedding model for 100+ languages: a 250,000-token lookup table, 256 dimensions, top MMTEB score among the multilingual static models Minish compared
minishlab/blink-multilingual-64MDeploy blink-multilingual-64M
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
blink-multilingual-64M is a multilingual static text embedding model from Minish, released on Hugging Face on November 17, 2026 under the MIT license. Like the English blink-8M, it has no transformer layers: each of its 250,000 tokens maps to a stored 256-dimension vector, and a text's embedding is the mean of its tokens' vectors. That makes it 64M parameters and 256 MB in float32, and it runs on a CPU at lookup speed in more than 100 languages.
Minish reports a mean of 51.60 across all 131 tasks of MTEB Multilingual v2, against 48.37 for bekko-static-embedding-v1 at the same 256 dimensions and 47.67 for its own potion-multilingual-128M, and above bekko's full 1024-dimension score of 50.18. It leads on bitext mining (57.03) and semantic similarity (67.47); on retrieval it is level with bekko (47.52 against 47.62).
It reads words without context, so it suits first-pass retrieval, deduplication and clustering of multilingual text, with a transformer embedder or reranker for the final ranking.
Architecture
A single 250,000 x 256 token embedding table (64.0M parameters) with mean pooling over the tokens of the input. No attention and no active layers: encoding is a tokenizer pass, a table lookup and an average.
Mixpeek SDK Integration
# Embed text on your side with model2vec, then upsert the vectors into a
# Mixpeek collection with a 256-d vector index. blink-multilingual-64M is not a
# managed extractor, so the inference runs where you choose.
import numpy as np, requests
from model2vec import StaticModel
model = StaticModel.from_pretrained("minishlab/blink-multilingual-64M")
texts = ["Forklift reverses into a shelf at dock 4", "La carretilla choca contra un estante"]
vecs = model.encode(texts) # shape (2, 256)
vecs = vecs / np.linalg.norm(vecs, axis=1, keepdims=True) # normalize for cosine
requests.post(
"https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
headers={"Authorization": "Bearer API_KEY"},
json={"collection_id": "col_your_collection",
"documents": [{"document_id": f"note-{i}",
"vectors": {"blinkml": v.tolist()},
"payload": {"text": t}}
for i, (t, v) in enumerate(zip(texts, vecs))]},
)Capabilities
- 256-dimension text embeddings in more than 100 languages from one 256 MB model
- No GPU needed: a static lookup and mean pool, with no transformer layers to run
- Loads with model2vec (Python), model2vec-rs (Rust) or Sentence Transformers
- MTEB Multilingual v2 mean 51.60 across 131 tasks, the best of the static models Minish compared
- MIT license, so commercial use is allowed
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| MTEB Multilingual v2 (131 tasks) | Mean(Task) | 51.60 | Model card: minishlab/blink-multilingual-64M (self-reported, 256d; bekko-static-embedding-v1 at 256d: 48.37) |
| MTEB Multilingual v2, retrieval tasks | Mean | 47.52 | Model card (self-reported; bekko at 256d: 47.62) |
| MTEB Multilingual v2, bitext mining | Mean | 57.03 | Model card (self-reported; best other static model: 48.50) |
| MTEB Multilingual v2, STS | Mean | 67.47 | Model card (self-reported) |
Performance
256 MB of float32 weights. Minish L2-normalizes every model's embeddings in its comparison, so normalize before cosine search. Speed figures are not published on the card, and we have not measured it.
Common Pipeline Companions
Frequently Asked Questions
What is blink-multilingual-64M?
A static multilingual text embedding model from Minish, released November 17, 2026 under MIT. It stores one 256-dimension vector per token in a 250,000-token table and averages them, so it has 64M parameters, weighs 256 MB and runs without a GPU in more than 100 languages.
How good is blink-multilingual-64M?
Minish reports a mean of 51.60 across MTEB Multilingual v2's 131 tasks, the best of the multilingual static models it compares (bekko-static-embedding-v1: 48.37 at 256 dimensions, potion-multilingual-128M: 47.67). Transformer embedders still score higher; use it where speed and size decide.
What is the difference between blink-8M and blink-multilingual-64M?
blink-8M is English-only, with a 30,000-token table and 31 MB of weights. blink-multilingual-64M covers more than 100 languages with a 250,000-token table and 256 MB. Both output 256-dimension vectors from a lookup and an average.
Can blink-multilingual-64M match text across languages?
Yes. It scores 57.03 on MMTEB's bitext mining tasks, which find translations of a sentence in another language, against at most 48.50 for the other static models Minish compared.
Can I use blink-multilingual-64M commercially?
Yes. It is released under the MIT license.
How do I use blink-multilingual-64M with Mixpeek?
Run the model on your side with model2vec or Sentence Transformers, normalize the 256-dimension vectors, and upsert them into a Mixpeek collection with a 256-d vector index. A retriever then searches them with a query embedded by the same model, in any of its languages.
Specification
Research Paper
blink-multilingual-64M model card (Minish, November 2026)
arxiv.orgBuild a pipeline with blink-multilingual-64M
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free