Technical blog

Booting Fast and Slow

At H Company, our agents are served by a fleet of vLLM replicas running on GPU nodes in Kubernetes.

Read the article

Illustration of one ice-covered GPU beside one glowing, flame-covered GPU.

Latest stories

Illustration of a small glowing ghost rising from a graphics card on a desk, next to a potted plant.

June 25, 2026Engineering

VRAM Ghost Busting: Who You Gonna `close()`?

At H Company, our training workloads run on GPU clusters managed by SkyPilot across Kubernetes and Slurm backends.

Line illustration of images, documents, and video feeding into a model that outputs a network of connected nodes.

June 1, 2026Research

Holo3.1 makes computer use fast and local

Line illustration of images, documents, and video feeding into a model that outputs a network of connected nodes.

April 28, 2026Research

Holotron3

All articles

Research releases and engineering notes, newest first. Filter by topic or search the index.

Booting Fast and Slow

VRAM Ghost Busting: Who You Gonna `close()`?

Holo3.1 makes computer use fast and local

Holotron3

Holo3 breaks the computer use frontier

Unlocking Online RL and Unifying Our AI Platform Across Kubernetes and Slurm

Holotron-12B

February 27, 2026

Computer-use for GEO

February 3, 2026

Holo2 235B

November 13, 2025

Holo2: Cost-Efficient Models for Cross-Platform Computer-use Agents