{ "id": "multilingual-e5-large-instruct", "identity": { "family": "e5", "display_name": "Multilingual E5 Large Instruct", "hub_id": "intfloat/multilingual-e5-large-instruct", "api_name": null, "revision": null, "variants": [], "license": "MIT", "commercial_use": true, "commercial_use_note": "MIT weights and code; no use restrictions.", "release_date": "2024-02", "paper_url": "https://arxiv.org/abs/2402.05672", "repo_url": "https://github.com/microsoft/unilm/tree/master/e5" }, "task": { "category": "embedding", "modalities_in": [ "text" ], "modalities_out": [ "vector" ], "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." }, "io": { "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" } } }, "mixpeek_wiring": { "extractors": [ { "extractor": "text_extractor@v1", "selected_by": "embedding_model unset or multilingual_e5_large_instruct_v1 (registry TEXT default)" }, { "extractor": "multimodal_extractor@v1", "selected_by": "fixed (transcription embedding)" }, { "extractor": "multimodal_extractor@v2", "selected_by": "fixed (transcription; OCR and description embeddings when those outputs run)" }, { "extractor": "document_graph_extractor@v1", "selected_by": "fixed" }, { "extractor": "web_scraper@v1", "selected_by": "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_embedding", "multimodal_extractor_v1_transcription_embedding", "multimodal_extractor_v2_transcription_embedding", "multimodal_extractor_v2_ocr_embedding", "multimodal_extractor_v2_description_embedding", "document_graph_extractor_v1_text_embedding", "intfloat__multilingual_e5_large_instruct" ], "retriever_stages": [ "feature_search", "query_expand", "mmr" ], "composes_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)" ] }, "run": { "status": "served", "status_reason": "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_format": "safetensors", "trust_remote_code": false, "size_on_disk_mb": 1120, "params_millions": 560, "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_shape": "both", "batch_size": 64, "serve_batch": true, "scale_to_zero": { "min_replicas": { "expr": "int(os.environ.get('MVS_E5_MIN_REPLICAS', '0'))" }, "max_replicas": { "expr": "int(os.environ.get('MVS_E5_MAX_REPLICAS', '10'))" }, "idle_timeout_s": 300, "memory_gb": 2, "num_cpus": 0.5, "num_gpus": { "expr": "hardware_config.gpu_fraction if hardware_config.gpu_concurrency > 0 else 0" }, "cold_start_s": { "warm_node_s": 31, "new_node_s": 365, "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" }, "cpu_fallback": true } }, "cost": { "unmeasured": true, "measured": null, "note": "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." }, "evidence": { "unmeasured": false, "items": [ { "dataset": "LongMemEval subset, 60 questions (hybrid_flat on Mixpeek)", "metric": "retrieval hit rate, SYSTEM number: E5 dense + BM25 fused by RRF in one retriever, top 20", "k": 20, "score": 1, "baseline": null, "baseline_score": null, "held_out": false, "date": "2026-09-27", "source": "research/agent-memory/RESULTS.md" }, { "dataset": "LongMemEval subset, 60 questions (hybrid_flat on Mixpeek)", "metric": "QA accuracy, SYSTEM number: E5 + BM25 RRF retrieval, Gemini v2 reader and Gemini judge", "k": null, "score": 0.683, "baseline": "local MiniLM dense + rank_bm25 RRF, same reader", "baseline_score": 0.783, "held_out": false, "date": "2026-09-27", "source": "research/agent-memory/RESULTS.md" } ] }, "refs": [ "hf://intfloat/multilingual-e5-large-instruct@274baa43b0e13e37fafa6428dbc7938e62e5c439" ], "formats": [ { "ref": "hf://intfloat/multilingual-e5-large-instruct@274baa43b0e13e37fafa6428dbc7938e62e5c439", "format": "safetensors", "file": "model.safetensors", "loader": "transformers / open_clip from_pretrained; Mixpeek Hub import (api/models/huggingface.py)", "supported_today": true }, { "ref": "hf://intfloat/multilingual-e5-large-instruct@274baa43b0e13e37fafa6428dbc7938e62e5c439", "format": "onnx", "file": "onnx/model.onnx", "loader": "onnxruntime; Mixpeek model upload (api/models/loadability.py, ONNX IR <= 11)", "supported_today": true } ], "use_cases": [ "text-search", "multilingual-search", "transcripts" ], "recommended": { "value": true, "for": [ "text-search", "multilingual-search" ], "reason": "Mixpeek's default text embedding: served at zero replicas, used by text, multimodal and document extractors, with committed retrieval evidence." }, "demos": [], "seo": { "title": "Multilingual E5 Large Instruct: embedding model", "description": "Multilingual E5 Large Instruct is a text embedding model (MIT). Served on Mixpeek, scaling to zero when idle.", "keywords": [ "Multilingual E5 Large Instruct", "e5", "embedding", "intfloat/multilingual-e5-large-instruct", "text-search", "multilingual-search", "transcripts" ], "faqs": [ { "q": "What is Multilingual E5 Large Instruct best at?", "a": "Default text vector for search over transcripts, OCR text, descriptions, documents and web pages in 100+ languages; one space for query and passage text." }, { "q": "What are the limits of Multilingual E5 Large Instruct?", "a": "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." }, { "q": "Can Multilingual E5 Large Instruct be used commercially?", "a": "Yes. MIT weights and code; no use restrictions." }, { "q": "How does Multilingual E5 Large Instruct run on Mixpeek?", "a": "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)." } ] }, "last_verified": "2026-11-05", "sources": [ "https://huggingface.co/intfloat/multilingual-e5-large-instruct", "https://huggingface.co/api/models/intfloat/multilingual-e5-large-instruct", "https://arxiv.org/abs/2402.05672", "server/engine/inference/intfloat/multilingual_e5_large_instruct/services.py", "server/engine/inference/intfloat/multilingual_e5_large_instruct/routes.py", "server/engine/extractors/text_extractor/v1/definition.py", "server/shared/models/embeddings/registry.py", "server/engine/extractors/multimodal_extractor/v2/definition.py", "server/engine/extractors/web_scraper/v1/definition.py", "server/engine/extractors/document_graph_extractor/v1/definition.py" ], "registry": { "intfloat__multilingual_e5_large_instruct": { "deployment": "intfloat__multilingual_e5_large_instruct", "service_id": "intfloat/multilingual-e5-large-instruct", "routes": "engine/inference/intfloat/multilingual_e5_large_instruct/routes.py", "route_prefix": "/multilingual_e5_large_instruct/v1", "enabled": true, "description": "Multilingual E5 Large Instruct embedding model. Generates 1024-dimensional dense embeddings for text. Supports 100+ languages with instruction-tuned representations.", "capabilities": { "category": "embedding", "modality_input": [ "text" ], "modality_output": "vector", "supports_batch": true, "reranking_mode": null }, "vector_config": { "dimensions": 1024, "distance": "cosine", "vector_type": "dense" }, "serve": { "min_replicas": { "expr": "int(os.environ.get('MVS_E5_MIN_REPLICAS', '0'))" }, "max_replicas": { "expr": "int(os.environ.get('MVS_E5_MAX_REPLICAS', '10'))" }, "idle_timeout_s": 300, "target_ongoing_requests": 1, "num_cpus": 0.5, "num_gpus": { "expr": "hardware_config.gpu_fraction if hardware_config.gpu_concurrency > 0 else 0" }, "memory_gb": 2, "pinnable": false, "placement": "worker", "enabled": true }, "interactive_twin": true, "interactive_serve": { "min_replicas": { "expr": "int(os.environ.get('MVS_E5_INTERACTIVE_MIN_REPLICAS', '0'))" }, "max_replicas": { "expr": "int(os.environ.get('MVS_E5_INTERACTIVE_MAX_REPLICAS', '2'))" }, "idle_timeout_s": 300, "target_ongoing_requests": 1, "num_cpus": 0.5, "num_gpus": { "expr": "hardware_config.gpu_fraction if hardware_config.gpu_concurrency > 0 else 0" }, "memory_gb": 2, "pinnable": false, "placement": "head", "enabled": true }, "schemas": { "input": "shared.inference.intfloat.multilingual_e5_large_instruct.models.InferenceRequest", "output": "shared.inference.intfloat.multilingual_e5_large_instruct.models.InferenceResponse", "config": "shared.inference.intfloat.multilingual_e5_large_instruct.models.InferenceConfigs" } } }, "templates": [ "contextual-video-match", "creative-dna", "ux-session-analysis", "video-moderation" ], "usage": { "namespaces": null, "collections": null, "uncounted": true, "templates": 4, "counted_at": null, "source": null }, "facets": { "status": "served", "category": "embedding", "family": "e5", "license": "MIT", "commercial_use": true, "modalities_in": [ "text" ], "modalities_out": [ "vector" ], "use_cases": [ "text-search", "multilingual-search", "transcripts" ], "recommended": true, "release_date": "2024-02", "params_millions": 560, "size_on_disk_mb": 1120, "vector_dims": 1024, "vector_distance": "cosine", "weights_format": "safetensors", "formats": [ "onnx", "safetensors" ], "loadable_today": true, "trust_remote_code": false, "cpu_fallback": true, "num_gpus": null, "memory_gb": 2, "min_replicas": null, "serving_shape": "both", "served": true, "ref_schemes": [ "hf" ], "has_evidence": true, "cost_measured": false, "template_count": 4, "usage_namespaces": null, "usage_collections": null, "usage_uncounted": true, "has_demo": false, "rank_score": 11 }, "rank_score": 11 }