Google Colab CLI launches GPU/TPU sessions

Google has shipped version 0.6.dev7+g510115b0c of the Colab command-line interface (CLI) [DevTo]. The tool is installable with pip install google-colab-cli or uv tool install google-colab-cli, and it immediately exposes a set of sub-commands for provisioning remote sessions, managing dependencies, and retrieving artifacts.

The Colab CLI supports NVIDIA T4, L4, A100, H100 GPUs and v5e1/v6e1 TPUs. Key features include:

• colab new [-s SESSION_NAME] [--gpu T4|L4|A100|H100] [--tpu v5e1|v6e1] to launch a fresh notebook instance
• colab install torch torchvision matplotlib to install pip packages inside the remote environment
• colab exec -f ./script.py to run a Python script on the remote VM, streaming stdout back to the terminal
• colab drivemount /mnt/drive to mount Google Drive for persistent storage
• colab console to drop the user into a Bash shell inside the running notebook

The CLI lacks the newer Blackwell G4 GPU tier (96 GB VRAM) available in the web UI, and sessions appear as “Unknown notebook” in the Colab web console – an issue tracked on GitHub [GitHub Issue].

By exposing Google’s cloud accelerators through familiar CLI primitives, the Colab CLI lowers the barrier for automation-heavy teams and positions Google’s notebook service as a programmable compute endpoint. The remaining hardware parity gaps will determine whether the CLI can become the default entry point for serious model training.