Pearl Network

What is Pearl Network?

Pearl Network is an AI inference platform by Pearl Research that provides access to open large language models through optimized serving infrastructure. It targets developers and organizations running model inference via API, with options ranging from serverless endpoints to on-premise deployments. Access requires an API key for the public endpoint or a scheduled call for managed and on-premise tiers.

What are the features of Pearl Network?

• OpenAI-compatible API

  Serverless endpoints accept requests using standard OpenAI client libraries pointed to the Pearl base URL.
• KV cache quantization (PearlQuant)

  Compresses the KV cache 65% below FP8 without losing quality or speed, according to the vendor.
• Supported models

  Serves models including deepseek-ai/DeepSeek-V4-Pro-0813, Qwen/Qwen3.8-27B, and zai-org/GLM-5.3 with context windows up to 1M tokens.
• Managed inference

  Offers reserved throughput, private routing, and support for custom or fine-tuned weights.
• On-premise deployment

  Installs the serving stack inside a customer's own environment so inference traffic does not leave their network.

What are the use cases of Pearl Network?

• Sustained production workloads

  Managed inference provides dedicated capacity and reserved throughput for continuous usage rather than variable serverless billing.
• Strict data residency requirements

  On-premise deployment keeps all inference traffic within the organization's own network and hardware.

How to use Pearl Network?

1. Set the API key

   Store your Pearl API key in the PEARL_API_KEY environment variable.
2. Configure the client

   Initialize the OpenAI Python client with base_url set to https://inference.pearlresearch.ai/v1.
3. Send a request

   Call client.chat.completions.create specifying a supported model identifier such as deepseek-ai/DeepSeek-V4-Pro-0813 and your message payload.

Pearl Network Pricing Information:

• Serverless

  Pay per token; no cluster reservations, idle GPU charges, or billing between requests. Specific per-token rates are not stated.
• Managed Inference

  Custom pricing for dedicated capacity; requires booking a call.
• On Premise

  Custom pricing for self-hosted deployment; requires booking a call.