novita.ai

Key features include:

• Serverless Model APIs: Access to 200+ models with low latency (200ms) and high uptime (99.5%), billed per token.
• Dedicated Endpoints: Private endpoints with guaranteed performance and no noisy neighbors, ensuring consistent latency.
• Agent Sandbox: Secure, isolated runtimes for AI agents to run tasks, use tools, and call models.
• GPU Instances: Full-control dedicated GPU machines (e.g., NVIDIA H200, H100) with predictable performance.
• Serverless GPU: Automatic resource allocation, scaling to zero when idle, pay only for execution.
• Bare Metal: Dedicated physical GPU clusters for large-scale workloads.

Novita AI is suitable for various AI workloads, including:

• Running AI models for text, image, audio, video, and vision tasks via API.
• Deploying and scaling AI agents that need secure, isolated runtimes.
• Training and fine-tuning models on dedicated GPU instances.
• Large-scale inference and training requiring bare metal performance.
• Building production AI applications with reliable infrastructure and cost efficiency.

Pricing details are available for model APIs, with examples such as Kimi K3 at $3/Mt input and $15/Mt output, and Deepseek V4 Flash at $0.14/Mt input and $0.28/Mt output. GPU instances and serverless GPU are billed per second or per execution, with no idle costs for serverless. Specific pricing for GPU instances and bare metal is not listed on the page, but the platform claims up to 50% cost savings compared to major cloud providers.