Hugging Face’s Rust-based stack for serving transformer language models—tight integration with the Hub, safetensors, and the broader HF ecosystem.
TGI is built for teams already standardizing on Hugging Face artifacts: pull a model, configure sharding and quantization flags, and expose a text-generation HTTP API. Trade-offs versus vLLM depend on workload shape, hardware, and which features you need (tensor parallelism, speculative decoding hooks, etc.).
Practical note
Compare both with your models and concurrency on real hardware; marketing throughput numbers rarely transfer one-to-one.