Features
Native PyTorch eager execution on TPUs
Fused Eager mode (50-100%+ speed gains)
Distributed training (DDP, FSDP)
Mixed precision training with FP8 on Ironwood TPUs
Integration with PyTorch Lightning, Hugging Face Transformers
Zero static graph compilation required
Scale to 100K+ chip clusters
Open-source backend (torch-xla) on GitHub
XLA compiler integration for optimized performance
Training and inference with vLLM on TPU
Model serving with vLLM unified backend (JAX & PyTorch)
Compatibility with existing PyTorch codebases
Supports Gemma 4 inference on vLLM TPU
Integration with MaxText for LLM training
Integration with Metrax metrics library (JAX)
Autonomous vulnerability scanning and PR generation
PR review for quality, security, and best practices
Legacy COBOL to Java migration with testing
Incident triage and root cause analysis
Agent Canvas: manage multiple agents in one workspace
Cloud agents run even when your machine is off
Schedule recurring tasks via cron or event triggers
Trigger workflows from GitHub, Slack, PagerDuty
Model-agnostic: supports Claude Code, Codex, Gemini CLI, MiniMax M2.7
Isolated sandbox execution on VM or cloud
Fine-grained access control and budget guardrails
Audit trails for enterprise compliance
Self-hosted deployment in your VPC
Open-source SDK for custom agent development
In-chat model switching with saved LLM profiles
