Run vision, speech, and text models on Bright Data’s global compute fleet. Submit raw video, audio, images, or documents — the platform preprocesses, distributes, and collects results at fleet scale. One API.
Submit your data to a catalog model or your own — the platform handles preprocessing, distribution across the fleet, and results collection. One API for batch pipelines, real-time inference, and model training.
Your data arrives in its original form. The platform extracts, segments, and normalizes it server-side — then fans it across the fleet. No preprocessing code to write or manage.
Consumer-device compute keeps costs low without trading throughput. Built-in reliability, open-vocabulary catalog models, and pay-per-task billing remove the infrastructure overhead from enrichment at scale.
Catalog models run out of the box — pick a use case, create a project, and submit your data.
Convert speech files — interviews, call recordings, podcasts, or broadcast audio — to text at scale. Supports multiple languages and speaker formats.
Detect and locate objects or brand logos in images. Fast bounding-box detection for large image datasets with low per-task cost.
Classify documents, product reviews, or support messages by topic or sentiment. Embed passages for downstream ranking or clustering.
Label images by category or screen content for sensitive material at scale. Open-vocabulary: use your own label set without retraining.
Embed images to detect near-duplicates or power a visual search index. High-quality embeddings for similarity ranking and clustering.
Embed text passages for semantic similarity ranking and retrieval. Power search, recommendation, and RAG pipelines over large text corpora.
Extract structured data from page images, tables, and complex layouts. Layout- and table-aware — handles scanned PDFs and multi-column documents.
Four steps from data to results — no infrastructure to manage.
Sign in at app.dcmp.brightdata.com, go to Settings → API Keys, and create a key with inference permissions.
Choose a catalog model — or upload your own — and set the compute tier. The project holds the model, tier, and pipeline configuration.
Point the API at a manifest URL, zip archive, S3 bucket, or Google Drive folder. You get a job ID back immediately — the platform handles the rest.
Poll the job endpoint or configure a webhook for completion events. Results are durable — access them any time after the job finishes.
Real-world data isn’t clean tensors. Submit raw files — the platform preprocesses server-side before inference ever touches a device.
FFmpeg audio extraction + Silero VAD utterance segmentation — one video file becomes many discrete utterances, each a separate inference task.
Pages rendered to images, type-classified, and dispatched one per task. Multi-column, scanned, and mixed-layout documents handled automatically.
Format normalization, silence trimming, and speech-activity segmentation — any audio codec in, clean utterance segments out.
Any format decoded, resized, and normalized to model-input tensors. Preprocessing chains validated at project creation, not at submit time.
Manifest URLs (JSONL), zip archives, S3 buckets (IAM role-based), and Google Drive folders — the platform fetches and fans out automatically.
Declare the input type at project creation — preprocessing chains resolve and validate immediately, so bad config is caught before any data moves.
Pay for completed tasks only — no reserved infrastructure, no idle charges. Volume pricing available; contact us for rates.
Runs on CPU-class consumer devices. Best for text embeddings, lightweight classifiers, and language identification.
Runs on GPU-capable consumer devices. Best for vision models, speech transcription, and image embedding pipelines.
Dedicated GPU pool with guaranteed capacity. Best for real-time inference, SLA-bound workloads, and large-model inference.
GET /api/v1/catalog.
POST /api/v1/models/upload-token,
give it a name, and use the returned model_id when creating a
project. Uploaded models are versioned, account-scoped, and one model can back
any number of projects.
auto tier balances CPU and GPU automatically to meet your time goal.
Talk to us about your use case. We’ll provision your account, walk through the API, and validate your first batch before you scale.
Enter your work email — if we recognize it, we’ll send you a one-time code.