World-class text embeddings · 512 dimensions
Similar.dev is engineered for the highest power-to-weight ratio in embeddings: production-grade semantic quality in 512 float32 dimensions, served through an OpenAI wire-format API you can call in under a minute. Smaller vectors, smaller indexes, smaller bills.
from openai import OpenAI client = OpenAI( api_key="sme_your_account_...", base_url="https://api.similar.dev/v1", ) response = client.embeddings.create( model="similar.dev-en-v1-medium-512", input=[ "Compact vectors for production retrieval.", "Same wire format. Smaller footprint.", ], ) vectors = [item.embedding for item in response.data] # 2 documents -> 2 x 512 float32 components
Model catalog
Every model returns 512 float32 components through the same request and response shape. Start on the English flagship, switch IDs as your corpus changes, never rewrite a client.
Our highest-capacity English document embedding. The default choice for retrieval, clustering, deduplication, and semantic search over English text.
A lighter English variant on the identical contract. Ideal for high-volume pipelines where throughput and cost dominate the decision.
One shared semantic space for English, Spanish, and Portuguese documents. Same meaning lands in nearby vectors across all three languages.
Why size matters
A retrieval system pays for embeddings twice: once at inference, then every hour the index stays in memory, on disk, and across replicas. At ten million documents, a 3072-dimensional float32 index is roughly 123 GB before any overhead. The same corpus at 512 dimensions is about 20 GB.
That gap is the difference between a vector database that fits on one node and one that needs a cluster. Similar.dev targets the point where quality stays high and the operational footprint stays small.
See pricingRaw vector storage only. Illustrative arithmetic, not a benchmark.
Fewer bytes per vector also means fewer bytes on the wire per query, faster index builds, and cheaper replicas in every region you serve.
We are running MTEB-aligned evaluation on the English flagship now. Until those numbers are published we will not show leaderboard positions, rankings, or quality superlatives. A model-specific internal multilingual evaluation for similar.dev-lang3-lite-512 is already public on the production site.
Published evidence
On a held-out internal multilingual retrieval benchmark spanning six language directions, top-1 retrieval ranged from 99.48% to 99.73% across 10,000 examples. Top-10 recall was 100% across those directions. This is an internal result, not MTEB or independent third-party validation. Applies to similar.dev-lang3-lite-512.
The magic sauce
Most embedding models spend their capacity evenly across a space that meaning does not fill evenly. We treat that imbalance as an opportunity.
Similar.dev models are built with a proprietary process we call latent-space arbitrage. The premise is that semantic geometry, like any market, has mispriced regions: directions that consume dimensions while contributing little to how documents actually separate in real workloads. Our process identifies where the signal is under-valued and where the capacity is over-spent, then rebalances the space so every one of 512 dimensions carries weight it has earned.
We will not publish the method. We will publish the results, in the open, on public benchmarks, as they land. What we can say plainly today: the outcome is embedding quality that competes with far larger vectors at a fraction of the storage, bandwidth, and compute.
Roughly one sixth of a 3072-d footprint, before quantization.
Less memory per shard, fewer shards per corpus, faster nearest-neighbor scans.
Capacity is reinvested where meaning concentrates, not spread thin.
Built to be depended on
Switching embedding models means re-embedding every document you own. We think you should never be forced into that by a deprecation notice.
Every model ID we publish stays callable for at least five years from launch before any retirement notice. Invest in a space knowing it will still be there.
Bring your own domain text and receive a compact embedding space calibrated to how your documents actually differ. Not yet available; join the interest list from the contact page.
Dedicated single-language models, each with its own ID and its own five-year commitment, rolling out beyond today's English and multilingual catalog.
Pricing
Developers and researchers start immediately with prepaid credits. Startup and Enterprise plans begin with a conversation so capacity, SLAs, and support are matched honestly.
Load a minimum $10 credit and start executing queries. Reload whenever you want; nothing recurring.
Higher volume, migration support, and early access to new model IDs, sized to your roadmap rather than a rate card.
SLAs, security review, and procurement-friendly billing agreed with your team. A direct intake path, not promises on a page.
Ready when you are
Sign up, load a $10 credit, and embed your first documents before your coffee cools.
base_url to https://api.similar.dev/v1.