{
  "data": null,
  "kind": "anchor.page",
  "links": {
    "api": "https://www.anchorterminal.com/api/v1/index.json",
    "html": "https://www.anchorterminal.com/niches/embeddings.model-migration/",
    "json": "https://www.anchorterminal.com/niches/embeddings.model-migration/index.json",
    "llms": "https://www.anchorterminal.com/llms.txt",
    "markdown": "https://www.anchorterminal.com/niches/embeddings.model-migration/index.md",
    "mcp": "https://www.anchorterminal.com/mcp",
    "slim": "https://www.anchorterminal.com/niches/embeddings.model-migration/index.min.md"
  },
  "markdown": "No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.\n\n| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |\n| --- | --- | --- | --- | --- |\n| 0 | 17 | 114 | 9 | 8 |\n\n## What counts as done\n\nThe listing publishes dated model ids and a deprecation policy with a notice period, says which models share a vector space, and returns the same vector for the same input on a pinned id and runtime; where an open-weight copy exists the docs say it reproduces the hosted vectors.\n\nMeasured, for our own tests: cosine similarity of 0.999 or more between two calls on the same pinned id.\n\nHow agents ask for it: our embedding model is being deprecated, what now; upgrade to the newer embedding model; will these vectors stay the same next year; move off this provider\n\n## Does the job, with a limit\n\nIt does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.\n\n- **Amazon Nova Multimodal Embeddings (Amazon Web Services, Inc.)**: A versioned id (amazon.nova-2-multimodal-embeddings-v1:0) and a lifecycle policy (EOL no sooner than 28 October 2026, legacy period of at least 6 months), but nothing on shared vector spaces or whether a pinned id returns the same vector.\n- **Cohere Embed and Rerank (Cohere Inc.)**: A deprecation policy with email notice and dated retirements, and Pro and Fast share a space; ids are versioned (embed-v5.0-pro) but not dated, and nothing says a pinned id returns the same vector.\n- **EmbeddingGemma 2 (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author))**: Open weights can be pinned by commit and kept indefinitely, which an account chose for that reason, but the card says nothing on reproducibility across runtimes, dtype or hardware, and nothing on sharing a space with EmbeddingGemma 300M.\n- **EmbeddingGemma 300M (Google LLC (RDAP registrant of gemma.ai; the card names Google DeepMind as author))**: Weights unchanged since 25 September 2025 and pinnable, but no statement on reproducibility across runtimes, quantised builds or hardware, or on any shared space with EmbeddingGemma 2.\n- **Gemini Embedding (Google LLC)**: Release and shutdown dates per model (001 shuts 14 May 2028), advance-notice policy, and 001 and 2 are documented as incompatible spaces; ids are undated aliases and nothing says a pinned id returns the same vector.\n- **Granite Embedding Small English R2 (International Business Machines Corporation)**: Open weights pinnable by commit, but no deprecation policy, no statement on reproducibility across runtimes, and the last assessment found model.safetensors replaced on 6 January 2026 without a note.\n- **Jina Embeddings and Reranker (Jina AI GmbH)**: The catalogue lists model ids and creation dates, but no deprecation policy, notice period or changelog was found, and nothing says a pinned id returns the same vector.\n- **jina-embeddings-v5-omni-small (Jina AI GmbH, a subsidiary of Elastic N.V. (per the jina-embeddings listing's 2 October 2026 check))**: The card says vectors share a space with jina-embeddings-v5-text-small, so a text index extends without re-embedding, but nothing says local vectors reproduce the hosted API's or stay the same across vLLM and Transformers.\n- **Mistral Embed and Codestral Embed (Mistral AI (RCS Paris 952 418 325))**: Dated ids (mistral-embed-2312, codestral-embed-2505) and a deprecation table with deprecation and retirement dates, but no stated notice period, no shared-space note and nothing on vector stability.\n- **Nomic Embed (Nomic, Inc.)**: The docs say local embeddings match the API's within a small margin of error, but there are no dated ids or deprecation policy, and the legacy reference warns endpoints may be removed.\n- **NVIDIA NeMo Retriever Embedding and Reranking NIMs (NVIDIA Corporation)**: Model ids name the precision (BF16, NVFP4), but the default checkpoint changes by GPU, FP16 on some and NVFP4 on others, so a pinned container doesn't guarantee the same vectors; release notes are undated.\n- **OpenAI embeddings (OpenAI OpCo, LLC)**: A deprecation policy with at least 6 months' notice for generally available models and email to active users, but ids are undated, no shared space is stated, and an account saw text-embedding-3-small listed for deprecation and then removed without explanation.\n- **Qwen3-Embedding-0.6B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: Open weights pinnable by commit, but an account and an open issue found vectors differ between vLLM versions and GPU models, with only identical hardware and version reproducing them.\n- **Qwen3-Embedding-4B (Alibaba Cloud (named in the copyright line of the 8B sibling's Apache 2.0 licence; this repository has no LICENSE file))**: Open weights pinnable by commit, but nothing in the card on reproducibility, and open issues report vectors differing across vLLM versions and across inference providers.\n- **Qwen3-Embedding-8B (Alibaba Cloud (named in the licence's copyright line))**: Open weights pinnable by commit, but nothing in the card on reproducibility, and open issues report vectors differing across vLLM versions and across inference providers.\n- **Qwen3-VL-Embedding-2B (Alibaba Cloud (RDAP names Alibaba Cloud (Singapore) Private Limited as the registrant of qwen.ai))**: Open weights pinnable by commit, but an account and a second user found vectors from vLLM 0.14 didn't match the offline Transformers path, so a pinned model isn't enough.\n- **Qwen3-VL-Embedding-8B (Alibaba Group (Tongyi Lab, named in the copyright line of the GitHub repository's Apache 2.0 licence))**: Open weights pinnable by commit, but nothing on reproducibility; the 2B sibling's vLLM vectors didn't match the Transformers path in an account.\n- **Voyage AI embeddings and rerankers (Voyage AI Innovations, Inc.)**: All 4-series models, including the open-weight nano, are documented as sharing a space, and older models stay accessible, but ids are undated, there is no deprecation policy with a notice period, and nothing on vector stability.\n\n- not this job: ZeroEntropy zerank and zembed (ZeroEntropy, Inc.): The hosted model was withdrawn and new signups closed, a named blocker; nothing says the open weights reproduce the hosted vectors, so an index built on the API needs re-embedding to be safe.\n\n## What stops agents finishing\n\n- hosted model withdrawn with endpoints answering 503 (ZeroEntropy) or listed for deprecation\n- runtime or GPU change alters vectors, so a pinned model is not enough\n- re-embedding cost (weeks of GPU for 200M vectors) with no adapter or shared space\n\n## Around it\n\nAfter: embeddings.bulk-backfill, vector-search.ingest-sync.\n\n## Where this comes from\n\nWeight 9 of 100 in Embeddings \u0026 rerankers: 8 of 62 accounts (91 account-use-case tags, accounts counted in each use case they fit, scaled to 100 and rounded by hand); OpenAI deprecation scare, Jina emergency migration, CLIP to SigLIP re-embedding cost and vLLM drift; ZeroEntropy's shutdown is in the context file, not an account\n\nAs data: https://www.anchorterminal.com/api/v1/use-cases/embeddings.model-migration.json, or the MCP tool get_use_case with id embeddings.model-migration. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.\n",
  "meta": {
    "attribution": "Anchor Terminal (https://www.anchorterminal.com)",
    "attributionRequired": true,
    "credit": "CC BY 4.0, credit Anchor Terminal and link the page or API address you took it from (https://www.anchorterminal.com/terms/#data)",
    "docs": "https://www.anchorterminal.com/docs/",
    "generatedAt": "2026-11-16",
    "license": "CC-BY-4.0",
    "method": "https://www.anchorterminal.com/benchmark/",
    "methodology": "0.8",
    "openapi": "https://www.anchorterminal.com/openapi.json",
    "preview": false,
    "run": "2026-10-01",
    "runLabel": "October 2026 research run",
    "terms": "https://www.anchorterminal.com/terms/#data"
  },
  "page": {
    "breadcrumbs": [
      {
        "name": "Home",
        "url": "https://www.anchorterminal.com/"
      },
      {
        "name": "For agents",
        "url": "https://www.anchorterminal.com/agents/"
      },
      {
        "name": "Niches",
        "url": "https://www.anchorterminal.com/niches/"
      },
      {
        "name": "Change or pin an embedding model without breaking the index",
        "url": ""
      }
    ],
    "description": "Change or pin an embedding model without breaking the index: 0 companies fully do it, 17 with a limit, 114 carry the tag. Every listing in Embeddings \u0026 rerankers with the reason and a source.",
    "facts": null,
    "h1": "Change or pin an embedding model without breaking the index",
    "image": "https://www.anchorterminal.com/assets/og/default.png",
    "path": "/niches/embeddings.model-migration/",
    "published": "",
    "section": "agents",
    "title": "Change or pin an embedding model without breaking the index | Anchor…",
    "toc": null,
    "updated": "2026-11-15",
    "url": "https://www.anchorterminal.com/niches/embeddings.model-migration/"
  },
  "tokens": {
    "markdown": 2000,
    "slim": 630
  },
  "version": 1
}
