<- all tokdocs
2026.08.23@future.with.ai98

NVIDIA Will Host China's Top AI Models for You, and an Open Source Agent Can Use Them

AI
Watch on TikTok

View on TikTok ->

NVIDIA's build.nvidia.com catalog serves Chinese frontier models like DeepSeek, GLM, Kimi, and MiniMax from NVIDIA's own infrastructure, and a free API key lets you drive them from Hermes Agent, an open source agent framework from Nous Research. The creator walks through the catalog on screen, points to where the key comes from, and pitches the combination as a no-cost agent stack. The core facts check out. The "free forever" framing deserves a closer look.

What NVIDIA Is Actually Offering

The screen recording shows the build.nvidia.com model catalog. NVIDIA hosts over 100 models on its own inference infrastructure through NIM (NVIDIA Inference Microservices) and exposes them through OpenAI-compatible API endpoints. Signing up for the free NVIDIA Developer Program requires only an email address, no credit card, and grants an API key with roughly 1,000 inference credits. The frames capture the catalog's breadth: NVIDIA's own Nemotron line (nemotron-3.5-lightning-30b-a3b, nemotron-3-ultra-550b-a55b), Meta's muse-glimmer-30b, translation and video super resolution models, and a banner for NemoClaw, NVIDIA's sandboxed agent execution product.

The Chinese Models in the Catalog

The video's hook is that NVIDIA runs Chinese labs' models on American supercomputers and hands out access. The catalog frames confirm the lineup: Z.ai's GLM-5.2 appears with an "agentic ai" tag, and DeepSeek's deepseek-v4-flash shows as a 284B MoE model with 13B active parameters aimed at coding, chat, and agentic workflows. The transcript also names Kimi (Moonshot AI) and MiniMax. These models sit alongside filters for third-party inference providers like OpenRouter, Together AI, and Deepinfra, so the same catalog doubles as a comparison point against paid hosts.

Hermes Agent as the Free Front End

The second half of the pitch is Hermes Agent, the open source agent framework from Nous Research (github.com/NousResearch/hermes-agent, MIT license). It runs in the terminal, as a desktop app, and inside messaging platforms and IDEs, and it works with any LLM provider, including OpenAI-compatible endpoints like NVIDIA's. Its distinguishing feature is a learning loop: when the agent solves a hard problem, it writes a reusable skill document and keeps persistent memory across sessions. The creator lists the use cases as building projects, researching competitors, and finding leads.

The Setup in Practice

The workflow shown is short. Create a free account at build.nvidia.com, generate an API key from the settings page (keys carry an nvapi- prefix), and point Hermes Agent at NVIDIA's endpoint with a supported model. Because the endpoints speak the OpenAI request format, any tool that accepts a custom base URL and key can use them the same way.

Where the Claims Stretch

Two numbers in the video did not survive verification. The creator says Hermes Agent has over 200,000 GitHub stars and calls it the fastest growing agent on the planet. Public sources describe the project at roughly 10,000-plus stars. He also says the models are "free forever," but NVIDIA's free tier is a credit grant of about 1,000 inference requests through the Developer Program, which is a trial allowance rather than a permanent commitment. The video ends with a link-in-bio pitch for the creator's own "Agent OS," so the numbers serve a funnel. The underlying setup is still real and costs nothing to try.

Key Takeaways

  • build.nvidia.com hosts 100+ models, including DeepSeek, GLM, Kimi, and MiniMax, on NVIDIA's own infrastructure.
  • A free NVIDIA Developer Program account grants an API key with about 1,000 inference credits, no credit card required.
  • The endpoints are OpenAI-compatible, so any tool that accepts a custom base URL can use them.
  • Hermes Agent (Nous Research, MIT license) is a solid open source front end for these models, with persistent memory and self-authored skills.
  • Treat the video's "200,000 stars" and "free forever" claims as marketing. The star count is closer to 10,000 and the free tier is a credit allowance.

Resources

Published August 23, 2026. Writeup generated from a favorited TikTok.