NEWVectors or files. Pick a path.Start →
    Served on MixpeekEmbeddingRecommendedVerified 2026-11-05

    Multilingual E5 Large Instruct

    intfloat/multilingual-e5-large-instruct

    Multilingual E5 Large Instruct is a text embedding model (MIT). Served on Mixpeek, scaling to zero when idle.

    A Ray Serve app in Mixpeek's inference registry hosts it. It starts at zero replicas and wakes on first use.

    Use in a new namespaceAdd to an existing namespace Record as JSON

    New opens the Designer in Studio. For an existing namespace, open it in Studio and use its Designer tab.

    No demo for Multilingual E5 Large Instruct yet

    A video, image, slideshow or playground goes here once we record one on our own data.

    What it is good and bad at

    Best at
    Default text vector for search over transcripts, OCR text, descriptions, documents and web pages in 100+ languages; one space for query and passage text.
    Worst at
    Exact keyword, id and number matches (pair it with BM25); text past 512 tokens is truncated; Mixpeek sends E5 prefixes (query: / passage:) rather than the model card's Instruct: template, so instruction tuning is unused.
    Task
    Embedding
    Input
    Text
    Output
    Vector
    Use cases
    Text search, Multilingual search, Transcripts
    Why we recommend it
    Mixpeek's default text embedding: served at zero replicas, used by text, multimodal and document extractors, with committed retrieval evidence.

    Templates that use it

    • Contextual video match

      An article goes in, a ranked playlist of video comes out, and every match carries the entities, keywords and categories that produced it. Both sides are embedded twice, once by a text model and once by a multimodal one, so nothing is embedded at query time and a clip in another language still matches.

      Deploy this template
    • Creative DNA for Ad Archives

      Search every scene of every ad you have run, and find the ones that keep earning spend

      Deploy this template
    • UX session analysis

      Moderated sessions become a corpus you can question. The transcript is joined to what was on screen at that second, so one query reaches every session at once and returns the moment rather than the recording.

      Deploy this template
    • Video Moderation Extraction Foundation

      Segment a video corpus, extract a keyframe record for every segment, and match brand marks against a reference set you upload. Face and public-figure matching run on a dedicated or customer-hosted deployment. This deploy is the extraction layer a moderation policy sits on, and the page carries a state on every part so you can see which policy pieces are yours to add.

      Deploy this template

    Input and output

    Input

    text
    max_tokens
    512
    truncation
    true
    languages
    100+
    prompts
    embedding_task retrieval_document -> 'passage: ', retrieval_query / semantic_similarity / classification / clustering -> 'query: ', none -> no prefix

    Output

    vector
    dims
    1024
    distance
    cosine
    type
    dense

    How Mixpeek wires it

    Extractors
    • text_extractor@v1 embedding_model unset or multilingual_e5_large_instruct_v1 (registry TEXT default)
    • multimodal_extractor@v1 fixed (transcription embedding)
    • multimodal_extractor@v2 fixed (transcription; OCR and description embeddings when those outputs run)
    • document_graph_extractor@v1 fixed
    • web_scraper@v1 generate_text_embeddings=true (default true)
    Feature URIs
    • mixpeek://text_extractor@v1/multilingual_e5_large_instruct_v1
    • mixpeek://multimodal_extractor@v1/multilingual_e5_large_instruct_v1
    • mixpeek://multimodal_extractor@v2/multilingual_e5_large_instruct_v1
    • mixpeek://multimodal_extractor@v2/multilingual_e5_large_instruct_ocr_v1
    • mixpeek://multimodal_extractor@v2/multilingual_e5_large_instruct_description_v1
    • mixpeek://document_graph_extractor@v1/intfloat__multilingual_e5_large_instruct
    • mixpeek://web_scraper@v1/intfloat__multilingual_e5_large_instruct
    Vector indexes
    text_extractor_v1_embeddingmultimodal_extractor_v1_transcription_embeddingmultimodal_extractor_v2_transcription_embeddingmultimodal_extractor_v2_ocr_embeddingmultimodal_extractor_v2_description_embeddingdocument_graph_extractor_v1_text_embeddingintfloat__multilingual_e5_large_instruct
    Retriever stages
    feature_search, query_expand, mmr
    Combines with
    • whisper transcript per video segment -> E5 transcription vector beside the Vertex or Gemini multimodal vector, fused by RRF/DBSF
    • E5 dense + BM25 full-text on the same text_extractor feature, fused by RRF
    • feature_search -> rerank (bge-reranker-v2-m3)

    How it runs

    Status
    Served on Mixpeek Ray Serve app in the inference registry, with an interactive twin for query embeds. Batch calls the Serve app when HARDWARE_E5_USE_API is true (the default); false loads E5 in the job (E5TextEmbeddingBatch).
    Serve apps
    intfloat__multilingual_e5_large_instruct
    Runtime
    transformers AutoModel + AutoTokenizer (XLMRobertaModel), average pooling
    Weights
    safetensors, no remote code
    Parameters
    560 M
    Size on disk
    1,120 MB
    Preprocessing
    Task prefix by embedding_task, tokenize with truncation at 512 tokens.
    Postprocessing
    Mean pool last hidden state over the attention mask, then L2-normalize (1024 dims).
    Serving
    both, batch 64, batched by Serve
    Hardware
    Runs on CPU (slower without a GPU)

    Scale to zero

    Minimum replicas
    0 by default (env MVS_E5_MIN_REPLICAS)
    Maximum replicas
    10 by default (env MVS_E5_MAX_REPLICAS)
    Idle timeout
    300 s
    Memory
    2 GB
    CPUs
    0.5
    GPUs
    Set at deploy time: hardware_config.gpu_fraction if hardware_config.gpu_concurrency > 0 else 0
    Cold start, node warm
    31 s
    Cold start, new node
    365 s
    Cold start source
    server/engine/inference/intfloat/multilingual_e5_large_instruct/routes.py interactive_serve note (2026-09-30): cold start with a node present 31 s; one that adds a spot node and pulls the image about 365 s

    Measured cost

    Unmeasured

    No committed throughput+latency+memory+cost measurement. Committed partial facts: interactive twin query embed p50 1.69 s, p90 1.98 s over 281 calls (services.py comment); batch defaults 64 texts on GPU, 16 on CPU.

    Our evaluations

    Dataset and metricScoreBaselineDate
    LongMemEval subset, 60 questions (hybrid_flat on Mixpeek)
    retrieval hit rate, SYSTEM number: E5 dense + BM25 fused by RRF in one retriever, top 20 (k=20)
    research/agent-memory/RESULTS.md
    		1None2026-09-27
    LongMemEval subset, 60 questions (hybrid_flat on Mixpeek)
    QA accuracy, SYSTEM number: E5 + BM25 RRF retrieval, Gemini v2 reader and Gemini judge
    research/agent-memory/RESULTS.md
    		0.683local MiniLM dense + rank_bm25 RRF, same reader: 0.7832026-09-27

    Questions

    What is Multilingual E5 Large Instruct best at?
    Default text vector for search over transcripts, OCR text, descriptions, documents and web pages in 100+ languages; one space for query and passage text.
    What are the limits of Multilingual E5 Large Instruct?
    Exact keyword, id and number matches (pair it with BM25); text past 512 tokens is truncated; Mixpeek sends E5 prefixes (query: / passage:) rather than the model card's Instruct: template, so instruction tuning is unused.
    Can Multilingual E5 Large Instruct be used commercially?
    Yes. MIT weights and code; no use restrictions.
    How does Multilingual E5 Large Instruct run on Mixpeek?
    Ray Serve app in the inference registry, with an interactive twin for query embeds. Batch calls the Serve app when HARDWARE_E5_USE_API is true (the default); false loads E5 in the job (E5TextEmbeddingBatch).