Last updated:

    Model Inference Overview

    What is Model Inference

    The platform provides one-click inference functionality, helping users quickly allocate compute and start inference services on supported model pages — no complex environment configuration needed. Under Model Inference there are two entry points:

    Entry Description
    Public Inference Shared services provisioned by administrators. Includes in-cluster local deployments and cloud API models forwarded through a unified gateway
    Dedicated Instances Exclusive deployments for a model you choose, with specification, framework, replica, and quantization settings

    Core Advantages

    • Flexible Invocation: Provides an intuitive web interface for conversation testing, while also generating standard API interfaces for business code integration.
    • Rich Framework Support: Supports multiple mainstream inference frameworks including vLLM, llama.cpp, SGLang, and TGI.
    • Instantly Available: Eliminates complex configuration by automatically launching container environments with all required dependencies.
    • VRAM guidance: The dedicated-instance create page shows Recommended Minimum GPU Memory and compares it with each specification.
    • Dual access: Public inference covers both local deployments and cloud APIs through one OpenAI-compatible interface.

    Supported Inference Frameworks

    Framework Features Best For
    vLLM High throughput, low latency, supports continuous batching Production-grade high-concurrency inference services
    SGLang Optimized for structured generation, supports RadixAttention Complex reasoning and structured output scenarios
    TGI (Text Generation Inference) Hugging Face's official inference server Inference within the Hugging Face ecosystem
    llama.cpp Supports GGUF format, runs on both CPU and GPU Resource-constrained environments or GGUF format models

    Inference Task Types

    The platform supports multiple inference task types — refer to the corresponding documentation for API usage: