# Embed v1 0.6b

Embed v1 0.6b is the low-cost option in Perplexity's pplx-embed-v1 family. It returns 1024-dimensional embeddings over inputs up to 32K tokens, with the same prompt-free integration and native quantization as the 4B tier.

- **Model ID:** `perplexity/pplx-embed-v1-0.6b`
- **Type:** embedding
- **Providers:** perplexity
- **Context window:** 32,000
- **Maximum output tokens:** 0
- **Pricing:** $0.004/1M input tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/pplx-embed-v1-0.6b

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { embed } from 'ai';

const result = await embed({
  model: 'perplexity/pplx-embed-v1-0.6b',
  value: 'Sunny day at the beach',
})
```

## About

Embed v1 0.6b is the smaller of Perplexity's two standard embedding models, released February 26, 2026. It targets high-volume embedding, where per-token cost rather than the last point of recall sets the ceiling on how much you can index. Like the 4B tier, it uses bidirectional attention with mean pooling and returns unnormalized INT8 vectors by default.

The size difference shows up in two places. Embed v1 0.6b returns 1024-dimensional vectors against 2560 for `pplx-embed-v1-4b`, so each vector takes less room in the index. Binary quantization is also less forgiving at this scale: Perplexity measures a 2 to 4 percentage point quality drop from binary output at 0.6B, against under 1.6 points at 4B. Test binary output on your own queries before you binarize a large index.

Retrieval quality holds up better than the parameter count suggests. On Perplexity's PPLXQuery2Query benchmark over a 2.4 million document corpus, Embed v1 0.6b reaches 71.1% Recall@10, against 61.8% for BGE-M3 and 55.1% for Qwen3-Embedding-0.6B. On the BERGEN end-to-end retrieval-augmented generation benchmark, Embed v1 0.6b outperforms the larger Qwen3-Embedding-4B on three of five question-answering tasks. Perplexity also reports that Embed v1 0.6b outperforms its Qwen3 counterpart at the same parameter scale on MTEB(Multilingual, v2).

## What to consider

Embed v1 0.6b returns unnormalized embeddings. Compare INT8 vectors with cosine similarity and binary vectors with Hamming distance. If your vector database only supports inner product, convert to float32 and L2-normalize before storing.

Binary quantization costs more quality here than on the 4B tier, 2 to 4 percentage points against under 1.6. Start with INT8 at the full 1024 dimensions and compress only when storage forces the issue.

Perplexity recommends embedding documents and queries with the same model. Mixing Embed v1 0.6b with `pplx-embed-v1-4b` across the two sides makes similarity scores unreliable, and moving between the two later means re-embedding the corpus.

## When to use

### Best For

- **High-volume embedding** where per-token cost sets the practical ceiling on corpus size
- **Compact vector indices** built on 1024-dimensional output
- **Retrieval prototyping** before you commit to a production model and dimension count
- **Multilingual semantic search** at the lower end of the pplx-embed cost range
- **Prompt-free integration**, since no instruction prefix can drift between indexing and query time

### Consider Alternatives When

- **Retrieval accuracy is the binding constraint**; `pplx-embed-v1-4b` returns 2560-dimensional vectors
- **You plan to binarize a large index**; the quality drop is larger at this parameter scale
- **Your corpus is exclusively source code**; `voyage-code-3` is purpose-built for that domain
- **You need generated text**; this model returns vectors only, not completions

## Best for

- **High-Volume Embedding:** Per-token cost sets the practical ceiling on corpus size
- **Compact Vector Indices:** 1024-dimensional output keeps the index small
- **Retrieval Prototyping:** Iterate on chunking and queries before locking a dimension count
- **Multilingual Semantic Search:** Evaluated on MTEB(Multilingual, v2) alongside the 4B tier
- **Prompt-Free Integration:** No instruction prefix can drift between indexing and query time

## Consider alternatives

- **Accuracy Is Binding:** `pplx-embed-v1-4b` returns 2560-dimensional vectors for higher recall
- **Large Binary Indices:** Binary quantization costs more quality at this parameter scale
- **Code-Only Corpora:** `voyage-code-3` is purpose-built for source code retrieval
- **Generated Text Required:** This model returns vectors only, not completions

## Frequently asked questions

### What embedding dimensions does Embed v1 0.6b produce?

1024 by default. Matryoshka representation learning lets you request shorter vectors through the `dimensions` parameter, trading a small amount of retrieval quality for less storage.

### How does Embed v1 0.6b compare to pplx-embed-v1-4b?

Embed v1 0.6b returns 1024-dimensional vectors against 2560 for `pplx-embed-v1-4b`, at a lower per-token price. The 4B tier reaches higher recall on Perplexity's published benchmarks and holds up better under binary quantization. Both share a context window of 32K tokens and need no instruction prefix.

### How does Embed v1 0.6b compare to Qwen3-Embedding-0.6B?

On Perplexity's PPLXQuery2Query benchmark over a 2.4 million document corpus, Embed v1 0.6b reaches 71.1% Recall@10 against 55.1% for Qwen3-Embedding-0.6B and 61.8% for BGE-M3. The `qwen3-embedding-0.6b` model is also available through AI Gateway if you want to compare on your own corpus.

### Can Embed v1 0.6b replace a larger embedding model?

Sometimes. On the BERGEN end-to-end retrieval-augmented generation benchmark, Embed v1 0.6b outperforms the larger Qwen3-Embedding-4B on three of five question-answering tasks. Test both on your own corpus before you decide.

### Which similarity metric should I use with Embed v1 0.6b?

Cosine similarity for INT8 embeddings and Hamming distance for binary embeddings. Embed v1 0.6b returns unnormalized vectors, so inner product and L2 distance give incorrect results unless you convert to float32 and L2-normalize first.

### Does Embed v1 0.6b need an instruction prefix?

No. You embed text directly, which removes the risk of a mismatch between indexing-time and query-time instructions.

### Should I use binary quantization with Embed v1 0.6b?

Test it first. Perplexity measures a 2 to 4 percentage point quality drop from binary output at the 0.6B scale, against under 1.6 points at 4B. INT8 is the default and already cuts storage 4x compared with FP32.

### What is the context window for Embed v1 0.6b?

32K tokens. Chunk documents longer than that before you embed them.

### How do I call Embed v1 0.6b through AI Gateway?

Use the AI SDK `embed` function for one value or `embedMany` for a batch, with the model identifier `perplexity/pplx-embed-v1-0.6b`. AI Gateway also serves embeddings through the OpenAI-compatible REST API.

### Is Zero Data Retention available for Embed v1 0.6b?

Zero Data Retention is not currently available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/pplx-embed-v1-0.6b)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://docs.perplexity.ai/docs/embeddings/standard-embeddings#models)
- [Provider pricing](https://docs.perplexity.ai/docs/getting-started/pricing#embeddings-api-pricing)


If you encounter incorrect information, can’t find what you need, need a capability Vercel doesn’t currently offer, or have any other feedback to report, please do it by following the instructions at https://vercel.com/feedback.md.