H

Forging Superintelligence. Putting AI to work for you, sparking the agentic AI revolution.

[H @hcompany_ai]

OFFICIALLY LAUNCHED
H's computer-use agent API now runs on desktop and web, built for devs and builders. Fully managed, powered by our #1 OSWorld-Verified model.

[H @hcompany_ai]

Cool kids build computer-use agents on saturdays

[H @hcompany_ai]

Coming soon

[Emrick @EmrickSini]

Just tried NVIDIA Snapshot on a 35B model on a single H100. Boot time went from 7.5 min (cold start from S3) to 46 s restoring from a snapshot 🏎️

Now testing TP>1, since v0.2.0 added support for it  ⏳

[H @hcompany_ai]

Online RL is only as fast as its inference.
At GTC Europe, @EmrickSini showed how we made RL steps 24% faster with elastic inference, fast replica booting and smart request routing.

[Emrick @EmrickSini]

We made online RL for computer-use agents 24% faster with 5% fewer GPU-hours.

At GTC Europe, I shared how we used prime-rl + NVIDIA Dynamo to post-train Nemotron Omni (30B-A3B) at H, with inference capacity that adapts during training🧵

[H @hcompany_ai]

@ArtAnosCranel @42AI_ @gladia_io @cognition @Linkup_platform @stripe Yes, this one's Paris only, but let us know which city you'd like us to come to next and we'll do our best!

[H @hcompany_ai]

As promised: after SF, Paris 🇫🇷

Computer Use Agents hackathon with @42AI_, November 7-8.

€10K cash prize, credits, more to be announced.
With @gladia_io, @cognition, @Linkup_platform and @stripe.

Apply here: https://t.co/hQRxv1wPJA

[H @hcompany_ai]

@Rainolf76160 Yes! On any build, click “Buy bricks”, downlaod the PDF instructions, paste the prompt into HoloTab, and order it!

[H @hcompany_ai]

Sidequest: we gave Holo4 a pile of bricks and a pile of blocks 🧱

Type an idea and watch it build.

HoloBricks makes brick models. HoloBlocks makes worlds you can walk around in.

Both built on our Agents API and open source.

Code and demos below ⬇️

[H @hcompany_ai]

Have a play: https://t.co/ASnzMCzc22

Both apps are open source, built with our TypeScript SDK and Agents API.

Fork them, add your own tools, share your weirdest build and tag us!

🧱 HoloBricks: https://t.co/GBMS8Za0uK
🌍 HoloBlocks: https://t.co/vutFIFhsxc

[H @hcompany_ai]

How it works: write code → run → look → revise.

Holo4 writes Python, runs it, checks renders from several angles, then makes changes.

Your browser’s GPU renders the build. A custom “look” tool sends the images back to the agent

[H @hcompany_ai]

Start with “a house held up by balloons.”

Then push it: bigger roof, twice the balloons, a dog on the porch.

You can steer while it builds, or select part of the model and ask for a change right there

[H @hcompany_ai]

Join the live session: https://t.co/cldjhbrIR5

[H @hcompany_ai]

Live with @NVIDIAAI this Thursday: how we serve vision-language models for computer-use agents.

Thu Oct 8 · 1PM CEST / 7AM ET
With @AvshalomM (inference team lead) and @EmrickSini (research engineer). Live Q&A.

[H @hcompany_ai]

SF Tech Week is next week. @tonywu_71 picked his favorites 👇

MongoDB hackathon, Boba & Build Night, no-cloud AI with ZETIC × Liquid, Physical AI Happy Hour, World Models & GPUs Coffee.

Oct 8: we're co-hosting the @skypilot_org AI Infra Meetup.

What are your picks?

[Tony Wu @tonywu_71]

I worked closely with Léonard on the SkyPilot training infra at H and I can promise you'll learn so much from his talk 🤝🏼

[H @hcompany_ai]

📍 San Francisco · Oct 8 · 5:30 PM
Part of @a16z Tech Week.
Limited seats: https://t.co/RyaVVo64Sg

[H @hcompany_ai]

Training computer-use agents means running thousands of environments next to the trainer, with sub-second cold starts so the GPUs never wait.
We do it on @skypilot_org.
On Oct 8, Leonard Benedetti, core researcher at H, will walk through how we train, how we eval, and what actually breaks when you scale up.

[H @hcompany_ai]

Finally, we are releasing the trajectories behind our public benchmark results so they can be inspected and replayed:
- Replay them at https://t.co/92MQ2QtnC4
- Download them from https://t.co/Bmc4bOfSOr

[H @hcompany_ai]

Today we're releasing Holo4, our new family of generalist computer-use models.

Holo4 can click, code, and call tools across desktop, web, Android, code sandboxes, MCP servers, and business APIs.

Available today on the H Models API for commercial use.

[H @hcompany_ai]

The training recipe combines supervised fine-tuning with two asynchronous online-RL experts, specialized for different classes of workflows and merged back into one generalist model.