Skip to content

Product And Model Lifecycle

Jump to bottom
Han Xiao edited this page Jul 2, 2026 · 3 revisions

The support, maintenance, and end-of-life (EOL) policy for self-managed (SM) / air-gapped Jina AI models. This page explains, in plain terms, how long a model you deploy will keep working, how long Jina/Elastic maintains it, what "end of life" actually means when the weights live inside your own network, and - importantly - what "maintenance" can and cannot mean for a machine-learning model as opposed to ordinary software. It is the reference for field, support, procurement, security, and audit questions about the lifetime of a Jina SM deployment.

Read this first. Hosted-API providers (OpenAI, Cohere, Azure Foundry, Amazon Bedrock) tie "end of life" to a shutdown date: after that date the endpoint is switched off and your requests fail. Air-gapped Jina SM is the opposite by design. You hold the weights inside your own perimeter, so a model you have deployed keeps working for as long as you choose to run it - there is no switch we can flip, no phone-home, no expiry. EOL here is about how long we keep maintaining and supporting a model, not about whether you can keep using it. If you take one thing from this page: deploying a Jina SM model is buying a permanent, self-contained artifact, not renting access to a service.

Contents

How SM lifecycle differs from a hosted API

flowchart LR
    classDef hosted fill:#ffe6e6,stroke:#c0392b
    classDef sm fill:#d9f5e0,stroke:#1f8f3a

    subgraph H[Hosted API - vendor controls the endpoint]
      H1[GA] --> H2[Legacy] --> H3[Deprecated] --> H4[Shutdown
inference STOPS - 410 Gone]:::hosted
    end

    subgraph S[Jina SM air-gap - you control the weights]
      S1[GA] --> S2[Legacy] --> S3[Maintenance] --> S4[End of Life
support ends - inference KEEPS RUNNING]:::sm
    end
Loading
Hosted API (OpenAI / Cohere / Azure) Jina SM (air-gap)
What EOL takes away Access to the model (endpoint shut down) Only our maintenance + support
Does inference stop at EOL? Yes - requests fail No - your bundle runs indefinitely
Who holds the weights The vendor You, baked into the bundle
Forced migration Yes, by the shutdown date Never - you migrate when it suits you
What ends the model's usefulness The vendor's calendar Your own quality / risk / compliance decision

Models are not software: what "maintenance" really means

This is the most common misunderstanding, and getting it wrong leads to unrealistic expectations. A model is not a program you can patch line by line. Its behaviour is baked into a fixed set of trained weights. That has concrete consequences for what maintenance can and cannot deliver.

What we CAN maintain during a model's supported life:

  • The runtime and server (server/app.py): API compatibility, new schema support, bug fixes in request handling, batching, and I/O.
  • Dependencies: patching a CVE in transformers, torch, or any bundled Python package by pinning to a fixed version and rebundling. See Versioning & Updates -> Catalog and dependency bumps.
  • Packaging: new base images, CPU/GPU build fixes, quantization variants of the same weights (for example an INT8/FP8 build to fit smaller hardware). Quantization is a lossless-to-near-lossless repackaging of the existing weights - it does not retrain them.
  • Documentation, sizing, and deployment guidance.

What maintenance CANNOT do - because it is physically not how models work:

  • We cannot "fix" a model's quality on a specific task by patching it. If a model underperforms on, say, a particular language or domain, there is no code change that repairs it. The only way to change what a model knows or how well it performs is to train new weights - and new weights are, by definition, a new model generation, not a patch to the old one.
  • We do not retrain or alter the weights of a released model. A shipped generation is frozen. This is a feature, not a limitation: it is exactly what guarantees that the same bundle produces bit-identical output forever (see determinism), which is what regulated customers require for reproducibility and audit.
  • A "quality upgrade" is always a migration to a newer generation, never an in-place fix. When we release, for example, a v5 text embedding model, that is the successor to v3 - a different model with different weights and different output vectors, not a repaired v3.

Rule of thumb. Treat a model version like a pressed vinyl record, not like a web app. We can fix the turntable (runtime), replace a worn needle (dependencies), and press the same recording onto a lighter disc (quantization). We cannot re-record the music without cutting a new record (a new model generation). Anyone expecting a model to be "hotfixed" on a task the way a software bug is patched is applying the wrong mental model - and will mis-plan their roadmap because of it.

This is why the lifecycle below separates runtime maintenance (which we do, on a schedule) from model quality (which only ever improves by adopting a new generation, on your timeline).

Lifecycle stages

Every Jina SM model generation is in exactly one stage. The stage describes the level of maintenance and support we commit to - never whether the model still runs for you.

Stage What it means New deployments Your existing deployment runs? Runtime maintenance (CVE / deps / bug fixes) Support (Q&A, guidance)
GA (General Availability) The current recommended generation for its capability. Recommended Yes Yes Full
Legacy A newer generation exists and is recommended, but this one is still fully supported. Migrate at your convenience. Allowed Yes Yes Full
Maintenance Superseded by a newer generation. Still receives security and critical runtime fixes; no new features. Migration recommended. Discouraged Yes Security / critical only Full
End of Life (EOL) We no longer issue fixes or provide standard support. The model keeps running in your environment - you own its continued operation, and you can keep running it as long as you like. Not recommended Yes (you operate it) None from us Best-effort / none

Note there is deliberately no "Shutdown" or "Retired" stage. Those stages only exist for hosted services with a vendor-controlled endpoint. In an air-gap deployment there is nothing for us to retire - the artifact is yours. See Versioning & Updates -> What a bundle pins and FAQ -> What if the model upstream gets pulled from HuggingFace?.

Maintenance term (how long we support a model)

The Maintenance Term is the window during which Jina/Elastic actively maintains a model generation - runtime security patches, critical bug fixes, dependency updates, and packaging support.

Because these are long-lived, on-premise and often air-gapped deployments in regulated environments - where re-certifying a new model is expensive and slow - we commit to a long horizon. For each Jina SM model generation, the Maintenance Term runs for at least the longer of:

  • 36 months (3 years) from the model generation's GA (initial public release) date, or
  • 24 months (2 years) from the GA date of its direct replacement generation,

whichever is later - and in practice we expect to maintain widely deployed generations for three to four years or more. As long as Jina/Elastic offers the air-gap product, we intend to keep maintaining the generations customers actually run in production; we will not strand a deployed customer on a short clock. This is deliberately longer than typical hosted-API windows (OpenAI commits to a 6-month minimum before retirement; Azure Foundry to an 18-month GA lifecycle) because those clocks govern shutting an endpoint off, which does not apply here - our clock only governs how long we keep shipping fixes.

Remember what "maintenance" covers here: runtime, dependencies, packaging, and quantization of the same weights - not model quality. Improved quality only ever arrives as a new generation, which you adopt on your own schedule (see Models are not software).

timeline
    title Maintenance Term for a model generation
    GA (T0) : Recommended, full support
    Replacement generation ships : Overlap window begins (>= 24 months of continued maintenance)
    T0 + 36mo OR replacement GA + 24mo (whichever later) : Maintenance Term ends
    After term : End of Life - keeps running for you indefinitely, no further fixes from us
Loading

Worked example

  • jina-embeddings-v3 GA: 2024-09-18
  • jina-embeddings-v5-text-* (its direct replacement) GA: 2026-02-18
  • Maintenance Term end = later of:
    • 2024-09-18 + 36 months = 2027-09-18
    • 2026-02-18 + 24 months = 2028-02-18 ← later
  • So jina-embeddings-v3 is maintained until at least 2028-02-18 - roughly two full years of overlap with v5 to evaluate, re-certify, and migrate at your own pace. And after that date, any v3 bundle you deployed still runs; it simply stops receiving new fixes from us.

Support term

Support (your questions, plus deployment / integration / sizing / migration guidance) tracks the parent Elastic support entitlement and applies while a model is in the GA, Legacy, or Maintenance stage. Formal support tiers, SLAs, and schedules are defined in the Jina SM Support Policy (in progress; see FAQ). For air-gapped and regulated deployments, support is delivered through your approved channel and never requires the model to reach the internet.

Notice and communication

Because nothing is ever force-retired, our notices are there to help you plan, not to warn you of impending loss of access.

Event Notice we give Where
New GA generation released (predecessor becomes Legacy) At release Model Catalog, release notes, models/catalog.json
A generation enters Maintenance At least 6 months before the Maintenance Term ends Release notes, this page, your account/support contact
A generation reaches End of Life Announced when the Maintenance Term ends This page, release notes
Security-critical issue that needs your action As soon as reasonably possible Support contact, security advisory

Determinism and migration guarantees

Licensing at EOL

Lifecycle stage and license are independent. Model weights stay under their original license (most current generations are CC-BY-NC-4.0; v1/v2 are Apache-2.0 - see Model Catalog). A CC-BY-NC-4.0 model in production still requires a valid Elastic commercial license at every lifecycle stage. Reaching EOL neither grants new license rights nor revokes your existing entitlement to keep running a licensed bundle.

Current lifecycle status

Generation-level status. Individual model IDs and per-model licenses are in the Model Catalog.

Capability GA (recommended) Legacy / Maintenance Notes
Text embedding jina-embeddings-v5-text-small / -nano (GA 2026-02-18) jina-embeddings-v3 (Maintenance, term to >= 2028-02-18) v2 base models (Apache-2.0) remain available for license-free evaluation
Multimodal embedding jina-embeddings-v5-omni-small / -nano (GA 2026-05-07) jina-embeddings-v4 (Legacy; Qwen Research License), jina-clip-v2 omni supersedes v4 multimodal
Code embedding jina-code-embeddings-1.5b / 0.5b (GA 2025-09-01) -
Reranker jina-reranker-v3 (GA 2025-10-01) jina-reranker-m0 (multimodal), jina-reranker-v2-base-multilingual
Reader / HTML-to-Markdown ReaderLM-v2 (GA 2025-01-16) reader-lm-1.5b / -0.5b
Vision-language (reader / OCR) jina-vlm (GA 2025-12-04) -

Dates are model GA (initial public release) dates from the official model catalog. This table is generation-level planning guidance; it is not a commitment to retire any model on a specific date, because SM models are not retired.

FAQ

If a model is at End of Life, do I have to stop using it? No. EOL only means we stop shipping fixes and standard support for it. Your deployed bundle keeps running exactly as before, indefinitely. Many air-gapped customers deliberately freeze a validated model for years.

Can you fix the model's accuracy on our specific documents/language without us migrating? No - and this is fundamental, not a policy choice. A model's quality is determined by its trained weights, which are frozen. There is no patch that changes what a model has learned. Better quality on a task comes only from a newer generation (new weights), which you adopt as a migration. What we can do within a generation is fix the runtime, patch dependencies, and provide quantized builds of the same weights.

How long will you really support the model we deploy today? At least 3 years from its GA, and at least 2 years after its replacement ships, whichever is later - and in practice we expect three to four years or more for widely deployed generations. As long as we offer this product, we maintain what customers run.

What happens to support if Jina/Elastic stops offering the product entirely? Your bundles still run - they are self-contained and yours. Because the toolkit, Dockerfiles, and CLI are Apache-2.0 and public, you (or a third party) can continue to re-bundle and operate any model you are licensed for from its weights. There is no license server or kill switch that could strand you.

Is quantization a "change" to the model? It is a repackaging of the same trained weights into a more compact numeric format, with quality held to a near-lossless bar. It is not retraining and does not alter what the model knows. We treat it as maintenance/packaging, not a new generation.

Next

Repo - Issues - Releases - Commercial license

Documentation source lives in docs/; edit there via PR and run ./scripts/sync-wiki.sh to mirror to the wiki.

Start here

Customer playbooks

Deploy

Reference

Contribute

Clone this wiki locally