AI Potluck
Contribute
Back to Gap Map Model components / Embeddings & retrieval models

Amazon Titan Text Embeddings

Amazon
closed / Overall score: 2.3

Amazon’s text embedding model on Bedrock, currently Titan Text Embeddings V2 (amazon.titan-embed-text-v2:0): an 8,000-token text embedder with configurable output dimensions defaulting to 1,024, invoked through the Bedrock runtime and used as the default embedder for Bedrock Knowledge Bases.

Still current but visibly aging. AWS’s model card records Titan Text Embeddings V2 with a 2024-04-30 launch and a lifecycle of Active, so it has not been superseded on paper; in practice Amazon now steers multimodal retrieval to the Nova line and this remains the text option. The Hugging Face repo amazon/Titan-text-embeddings-v2 is a card, not a release — it holds README.md and config.json and no weights — but its README carries a full 329-result MTEB model-index, which is the only published benchmark for the model. Verified 2026-09-11 via the Bedrock model card, the Bedrock pricing page and the Hugging Face model API.

Openness

1 high confidence
1.0
weights
closed(the amazon/Titan-text-embeddings-v2 Hub repo holds README.md and config.json only
the model runs solely on Bedrock as amazon.titan-embed-text-v2
0)
data
closed(training corpus not released or described)
code
closed(no training or inference implementation
license
proprietary(AWS end-user license over a Bedrock-hosted model
the Hub card records license
other and distributes nothing)

API-only, so the ladder's first rule fires. The Hub repo is the trap here and is recorded deliberately: amazon/Titan-text-embeddings-v2 looks like a weights release and is a model card with a config file, so the license field on it governs a README rather than a checkpoint.

  • https://huggingface.co/api/models/amazon/Titan-text-embeddings-v2 recorded 2026-09-11

    amazon/Titan-text-embeddings-v2: gated false, tags include 'license:other', and the file list is ['.gitattributes', 'README.md', 'config.json'] — no weight shards. Last modified 2024-04-30.

  • https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-titan-text-embeddings-v2.html recorded 2026-09-11

    The Bedrock model card gives model ID amazon.titan-embed-text-v2:0, an 8K context window, a 2024-04-30 launch date and 'Model lifecycle: Active'. Access is through the Bedrock runtime and the page offers no download. It is also the whole of Amazon's product documentation for this model, and it says nothing about the corpus - the strings "training data", "training corpus", "trained on" and "dataset" appear nowhere on it - which is what establishes data as closed rather than merely undocumented.

Adoption

3 medium confidence
3.0

No artifact to count — the Hub card's 518 monthly downloads measure a README, not a model — and no call volume published. The level rests on being a first-party model on Amazon Bedrock with its own list price and an Active lifecycle, and on its role as a default embedder for Bedrock Knowledge Bases. Distribution is broad but single-platform, which is why it sits below the models Microsoft and Google resell as well.

Capability

2 high confidence
2.0

Rung 2, narrow or dated. The evaluation is real and unusually complete for a closed model — 329 MTEB results published by the vendor — but it dates from the April 2024 release and nothing has replaced it. Short context by 2026 standards, text-only, and two generations behind the frontier while remaining useful for its niche, which is retrieval inside Bedrock. Level with mistral-embed: both are 8K-token text-only models of the 2023-24 generation still sold unchanged.

Verified 2026-09-11