FORGE

TURBO · FREE Your first vector in 60 seconds. Pipe a local corpus straight into GPU memory with forge-cli over gRPC and Protobuf — no HTTP hop.

playground.voxell.ai · no login required

voxell@playground:~$ forge embed --text \ "Can annual plans be refunded after renewal?"

0 msP50 · END TO END

200 embedding: [0.0123, -0.0456, 0.0789, …]

1024dturbo5 tokens

YOUR TURN · SAME CALL, YOUR CODE

curl https://api.voxell.ai/v1/embed \ -H "Authorization: Bearer $VOXELL_API_KEY" \ -d '{"texts":["Can annual plans be refunded after renewal?"],"model":"turbo"}'

Run it in the playground with no login, or paste it after you grab a key. Turbo is free. Prefer a terminal? forge-cli embed --text "..." is a public download →

87ms replayed at 1/12th speed · measured end to end, not time to first byte

THE GAP

FORGE87 ms

MARKET AVG300 ms

VQS74 / 100 · sample

nDCG@1070.0

TURBO · $0 outscores OpenAI's best embedding model VQS quality standard → > knowledge is precious

The gap, in practice

87ms vs 300ms. Same question, same corpus size. 213ms is the difference between an agent that chains ten calls and one that times out.

Deploy anywhere

Hosted at api.voxell.ai. Embedded in your VPC from the AWS Marketplace. Or on-prem as a Docker container behind your firewall. Same schema in all three.

Zero trust

mTLS client identity is available on every tier — Ed25519 client certificates in place of bearer tokens, no shared secrets to rotate. Every tier is encrypted in transit with TLS 1.3. Zero data retention after the response. Learn about VQS corpus quality scoring.

Three tiers

Turbo 1024d$0
Pro 2560dbalanced
Ultra 4096dmax fidelity

See pricing →

Numbers, with receipts: 87ms is P50 end to end on owned NVIDIA DGX hardware, no shared queue. 300ms is the observed market average across hosted embedding APIs. Turbo outscores OpenAI's best embedding model on English MTEB; the flagship holds rank #1 at 75.98 Mean (Task), independently verified on the public leaderboard. Methodology in the docs.