NEWVectors or files. Pick a path.Start →
    Models/minishlab/blink-8M
    MIT

    blink-8M

    by minishlab

    A 31 MB static English embedding model: a token lookup table with mean pooling, the top MTEB score among static models of its size

    17likes
    7.7M (30,000 x 256 embedding table)params
    Identifiers
    Model ID
    minishlab/blink-8M

    Deploy blink-8M

    Single-tenant

    Mixpeek 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-8M is a static English text embedding model from Minish, released on Hugging Face on November 8, 2026 under the MIT license. It has no transformer layers: each of its 30,000 tokens maps to a stored 256-dimension vector, and a text's embedding is the mean of its tokens' vectors. That makes it 7.7M parameters and 31 MB in float32, small enough to embed on a CPU, in a browser or inside a mobile app.

    Minish reports a mean of 55.35 on all 41 tasks of MTEB English v2, the highest of the seven static models it compares, and 38.86 on retrieval, 7.64 points above its own potion-base-8M at the same size. On the RTEB domain retrieval tasks (legal, finance, healthcare, code) it scores 30.84 nDCG@10, again the best static model in Minish's table. It replaces potion-base-8M.

    A static model trades accuracy for speed and size: it reads words without context, so it suits first-pass retrieval, deduplication, clustering and classification over large text sets, with a transformer embedder or a reranker for the final ranking.

    Architecture

    A single 30,000 x 256 token embedding table (7.7M parameters) with mean pooling over the tokens of the input. No attention and no active layers, so encoding is a tokenizer pass plus a table lookup and an average. Distilled with Minish's tokenlearn method from bge-base-en-v1.5 outputs on C4, then trained on LightOn's curated embedding pre-training and English fine-tuning sets.

    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-8M 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-8M")
    texts = ["Forklift reverses into a shelf at dock 4", "Pallet wrap tears during loading"]
    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": {"blink8m": v.tolist()},
                             "payload": {"text": t}}
                            for i, (t, v) in enumerate(zip(texts, vecs))]},
    )

    Capabilities

    • 256-dimension English text embeddings from a 31 MB model (float32)
    • 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 English v2 mean 55.35 and retrieval 38.86, the best of the static models Minish compared
    • MIT license, so commercial use is allowed

    Use Cases on Mixpeek

    First-pass retrieval over millions of short texts on CPU, with a reranker on the top results
    Deduplicating and clustering captions, transcripts and product titles before indexing
    Embedding text on device or in the browser where a transformer is too large
    Cheap text features for classification and routing in an ingestion pipeline

    Benchmarks

    DatasetMetricScoreSource
    MTEB English v2 (41 tasks)Mean(Task)55.35Model card: minishlab/blink-8M (self-reported, 256d; potion-base-8M: 53.30)
    MTEB English v2, retrieval tasksMean38.86Model card (self-reported; potion-base-8M: 31.22)
    MTEB English v2, reranking tasksMean43.12Model card (self-reported)
    RTEB English, 15 public tasks (legal, finance, healthcare, code)Mean nDCG@1030.84Model card (self-reported; best static model in the comparison, bekko-static-embedding-v1: 30.32)

    Performance

    Input SizeAny length of English text; every token is looked up and averaged
    Embedding Dim256
    GPU LatencyInput dependent
    GPU ThroughputBatch dependent
    GPU MemoryModel dependent

    31 MB of float32 weights. Minish's comparison evaluates every model at 256 dimensions and L2-normalizes the embeddings, which it reports improved every model, so normalize before cosine search. Speed figures are not published on the card.

    Frequently Asked Questions

    What is blink-8M?

    A static English text embedding model from Minish, released November 8, 2026 under MIT. It stores one 256-dimension vector per token in a 30,000-token table and averages them, so it has 7.7M parameters, weighs 31 MB and runs without a GPU.

    What is a static embedding model?

    A model whose embedding is a lookup and an average: each token has a fixed vector, and the text's vector is the mean of its tokens' vectors. With no transformer layers it is far faster and smaller, at the cost of reading words without their context.

    How good is blink-8M?

    Minish reports a mean of 55.35 across the 41 tasks of MTEB English v2 and 38.86 on its retrieval tasks, the best of the seven static models it compares and 7.64 retrieval points above potion-base-8M at the same size. Transformer embedders still score higher; use blink-8M where speed and size decide.

    Is blink-8M better than potion-base-8M?

    On Minish's numbers, yes on mean (55.35 against 53.30) and retrieval (38.86 against 31.22), and lower on classification (67.88 against 70.41). Both are 256-dimension models of the same size, and Minish says blink-8M supersedes potion-base-8M.

    Can I use blink-8M commercially?

    Yes. It is released under the MIT license.

    How do I use blink-8M 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.

    Specification

    Organizationminishlab
    Retriever-
    Parameters7.7M (30,000 x 256 embedding table)
    LicenseMIT
    Downloads/moN/A
    Likes17

    Research Paper

    blink-8M model card (Minish, November 2026)

    arxiv.org

    Build a pipeline with blink-8M

    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