Lightweight, efficient embedding model for high accuracy retrieval tasks. Supports flexible embedding sizes (1024, 512, 256) and 100+ languages.
Amazon Titan Text Embeddings V2 is Amazon's second-generation text embedding model optimized for Retrieval-Augmented Generation (RAG) workflows. Released in April 2024, it's designed to convert natural language text into high‑dimensional vectors with configurable output dimensions (256, 512, or 1024), enabling fine‑tuned trade‑offs between accuracy, storage cost, and retrieval latency.
As a managed embedding service within Amazon Bedrock, Titan V2 is fully serverless — no infra provisioning or runtime maintenance required. It supports over 100 languages in pre‑training, offers unit‑vector normalization for improved cosine‑similarity accuracy, and is priced at $0.02 per million input tokens, a 5‑fold reduction from the V1 model while typically maintaining >97% of its retrieval accuracy at 256 dimensions.
AWS serves Titan Text Embeddings V2 on the Bedrock bedrock-runtime endpoint with Invoke, text input, embedding output, an 8K-token / 50,000-character input ceiling, model id amazon.titan-embed-text-v2:0, and no cross-region inference profile on the current model card.
TheRouter serves the same embedding model as amazon/titan-embed-v2 through /v1/embeddings with OpenAI-compatible client syntax. Operational price, context, modality, and supported-parameter fields come from the live TheRouter catalog; the curated page documents AWS source-of-truth behavior and leaves snippets capped until an operator-funded verification run records execution.
| Type | Rate |
|---|---|
| Input | $0.0216 / 1M tokens |
| Release date | 2024‑04‑30Amazon Bedrock docs ↗ | verified |
| Output dimensions | 256, 512, 1024 (configurable)docs.aws.amazon.com ↗ | verified |
| Vector normalization | Yes (unit‑vector), optionaldocs.aws.amazon.com ↗ | verified |
| Supported languages | Optimized for English, with 100+ languages in preview; cross-language queries are sub-optimaldocs.aws.amazon.com ↗ | verified |
| Max input characters | 50,000docs.aws.amazon.com ↗ | verified |
| Training cutoff | Not publicly disclosed | unknown |
| Tokens‑to‑characters ratio (English) | ~4.7 chars/tokendocs.aws.amazon.com ↗ | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
MTEB (Massive Text Embedding Benchmark) weighted average | 60.37score | AWS ML blog ↗ | |
MTEB reranking | 57.5score | aws.amazon.com ↗ | |
MIRACL multilingual | 55score | aws.amazon.com ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "amazon/titan-embed-v2",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Generate embeddings for a text input with configurable dimensions and optional normalization. TheRouter routes your request transparently to Amazon Bedrock with fully OpenAI‑compatible syntax.
curl https://api.therouter.ai/v1/embeddings \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "amazon/titan-embed-v2",
"input": "Your text to embed here.",
"dimensions": 256,
"encoding_format": "float"
}'AWS announces the general availability of Titan V2 with configurable dimensions (256, 512, 1024), unit‑vector normalization, and 100+‑language support — priced at $0.02 per million tokens, a 5× reduction from V1.
AWS ML team publishes a comprehensive guide to Titan V2’s MTEB scores (weighted avg 60.37), multilingual performance, and RAG‑optimisation benefits.
Amazon measured that 512‑dimensional vectors retain ~99% of the retrieval accuracy of the full 1024‑dimension outputs, while 256‑dimension vectors keep >97% accuracy. This means you can reduce vector storage by 75% (256→1024) or 50% (512→1024) while sacrificing only marginal retrieval quality. For most RAG workloads, 256 dimensions strike an excellent cost‑accuracy balance.
While pre‑trained on 100+ languages, AWS explicitly notes that cross‑language queries will deliver 'sub‑optimal results'. The model excels at monolingual retrieval within its training languages; for robust cross‑lingual search, consider OpenAI text‑embedding‑3‑large or Cohere embed‑v4 which are designed for that case.
Embedding models process relatively short texts (≤8192 tokens) per request, so throughput is limited by per‑request computation overhead rather than token‑streaming capacity. RPM quotas ensure fair usage across tenants. For high‑volume indexing, request quota increases via AWS Support or use Bedrock Batch for throughput‑optimized bulk embedding.
Titan V2 introduces three major upgrades: (1) configurable dimensions (256/512/1024) vs G1’s fixed 1536; (2) unit‑vector normalization for improved cosine‑similarity accuracy, especially in RAG; (3) 100+‑language support (pre‑training) vs G1’s 25+. It's also 5× cheaper per‑token ($0.02 vs $0.10 per million) and retains >97% of G1’s retrieval accuracy at 256 dims.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | Amazon Bedrock docs ↗ | 2026-08-23 | verified |
| Output dimensions | docs.aws.amazon.com ↗ | 2026-08-23 | verified |
| Vector normalization | docs.aws.amazon.com ↗ | 2026-08-23 | verified |
| Supported languages | docs.aws.amazon.com ↗ | 2026-08-23 | verified |
| Max input characters | docs.aws.amazon.com ↗ | 2026-08-23 | verified |
| Training cutoff | — | — | unknown |
| Tokens‑to‑characters ratio (English) | docs.aws.amazon.com ↗ | 2026-08-23 | verified |
| MTEB (Massive Text Embedding Benchmark) weighted average | AWS ML blog ↗ | 2026-08-23 | verified |
| MTEB reranking | aws.amazon.com ↗ | 2026-08-23 | verified |
| MIRACL multilingual | aws.amazon.com ↗ | 2026-08-23 | verified |
| Amazon Titan Text Embeddings V2 now available in Amazon Bedrock | AWS News Blog ↗ | 2026-08-23 | verified |
| Get started with Amazon Titan Text Embeddings V2: A new state‑of‑the‑art embeddings model on Amazon Bedrock | AWS Machine Learning Blog ↗ | 2026-08-23 | verified |
| What are the practical accuracy‑storage trade‑offs between 256, 512, and 1024 dimensions? | aws.amazon.com ↗ | 2026-08-23 | to verify |
| Does Titan V2 support cross‑language search (e.g., query in German against a corpus in Korean)? | docs.aws.amazon.com ↗ | 2026-08-23 | to verify |
| Why does Bedrock throttle embedding models by RPM instead of TPM? What’s the performance implication? | docs.aws.amazon.com ↗ | 2026-08-23 | to verify |
| What’s the difference between Titan V2 and its predecessor Titan G1 (V1)? | aws.amazon.com ↗ | 2026-08-23 | to verify |