Skip to main content
Version: 0.15.x

Titan Takeoff

Titan Takeoff is a state-of-the-art inference server designed for self-hosting and deploying Large Language Models (LLMs).

It combines ease of use with efficient performance, enabling quick deployment of servers using a simple Docker command. This approach not only saves time but also optimizes server performance through techniques like quantisation and batching, leading to lower latency and higher throughput.

The server is particularly suited for users who need on-premise deployment for data privacy reasons or those who prefer control over their models. With Titan Takeoff, you can use fine-tuned and custom models in-house, without depending on external APIs. This feature is essential for industries where data sensitivity is a concern or where specialized model tuning is required.

For developers, Titan Takeoff reduces the time and effort needed to build and maintain the infrastructure for serving models, allowing them to focus on more critical aspects of their projects.

In essence, Titan Takeoff offers a straightforward, efficient way to deploy and manage Large Language Models while providing the flexibility and security of on-premise hosting.

Features:​

  • 🚀 A proprietary inference engine backend, offering best-in-class inference speed and throughput
  • 📦 Packaged in a single, easily deployed container, ready for self-hosted (and even offline) machines
  • 🎚️ Seamless multi-gpu and quantization support, with tools to find the best possible model for your hardware
  • 📡 Support for streamed responses, allowing easy design of interactive user applications
  • 🖥️ Handy GUI for testing out models and managing every aspect of Takeoff
  • 📥 Sophisticated batching behavior adapted to the task being performed
  • 🧑‍💻 Deployed with a single command, with inference then performed over a Rest API
  • 🎛️ Structured generation controls, ensuring outputs are legal JSON or adhere to a given regular expression
  • 🦀 Inference orchestrated by Rust for minimal infrastructure overhead
  • 🤗 Widespread support for huggingface models, plus support for custom models
  • 🏨 Support for hosting multiple copies of a model or multiple models, all from one instance
  • 🛠️ Model Management API for dynamically managing which models are served, with the ability to launch multiple models from a manifest file
  • 🤓 Dedicated team keeping on top of LLM-ops developments, ensuring neither Takeoff (nor its users) fall behind the curve
  • 🔌 Existing integrations with Langchain and Weaviate, with more planned
  • 📊 Metrics dashboard for monitoring usage
  • 📚 Deployment guides for AWS, GCP, Kubernetes, Vertex and Sagemaker
  • 💬 Monitored Discord server with support from both the team and the community