~80 open-weight models. One account. OpenAI-compatible. No bill.

NVIDIA has opened free API access to roughly 80 AI models through its build.nvidia.com catalog - including some of the most talked-about open-weight releases of the last six months: MiniMax M2.7 (230B parameters), GLM 5.1, Kimi 2.5, DeepSeek 3.2, and GPT-OSS-120B.

This isn't a sandbox or a demo. It's a production-grade API, OpenAI-compatible, with no rate limits and no expiring keys.

If you've been juggling subscriptions across OpenAI, Anthropic, Google, DeepSeek, and three open-weight providers, NVIDIA just collapsed that into a single account.

What you actually get

• ~80 models in the NIM catalog, refreshed regularly with new open-weight releases.
• Free API access at the time of writing. Free isn't forever, but it's free now.
• No request limits, no key expiration.
• Standard OpenAI-compatible API - drop-in replacement for most existing code.
• Production-ready - not a hosted notebook, an actual inference endpoint.

The headline model is MiniMax M2.7 at 230B parameters - the kind of model that, until last week, you could only run yourself on hardware most teams don't have, or rent at meaningful cost from a specialist inference provider.

How to get started (5 minutes)

1. Register. Go to build.nvidia.com and sign in with email or Google.
2. Verify. Confirm your email and phone (OTP). One-time only.
3. Get your API key. After completing your profile, generate a key. No request limits, no expiration.
4. Point your IDE at NVIDIA. In Cursor, Zed, OpenCode, Hermes, OpenClaw - or anything OpenAI-compatible - change the base URL to https://integrate.api.nvidia.com/v1 and paste your key.
5. Pick a model. Use a model ID like minimax-m2.7 or glm-5.1 and start sending requests.

That's it. Because the API is OpenAI-compatible, your existing code likely needs no changes beyond the base URL and key.

Why this matters strategically

NVIDIA sells GPUs. So why is it giving away inference?

Three reasons, and they all matter to anyone building on AI:

1. Developer mindshare. Whoever owns the default API endpoint owns the next generation of AI startups. OpenAI has it today. Anthropic and Google are competing for it. NVIDIA - sitting on the hardware everyone runs on - is now in that race too.
2. Up-the-stack expansion. NVIDIA is no longer just a chip company. With NIM, CUDA, Omniverse, and now a hosted API, it's becoming a full AI platform. The same playbook AWS ran on physical infrastructure twenty years ago.
3. Open-weight gravity. By hosting Chinese frontier models (MiniMax, GLM, Kimi, DeepSeek) alongside Western ones, NVIDIA is positioning itself as the neutral ground for open-weight inference. That's a meaningful pitch for enterprises that don't want to bet on a single lab.

What to watch for

• When does free end? Probably when the catalog gets too popular. Pricing is the obvious next step. Lock in your testing now.
• Latency and reliability at scale. Free APIs from large vendors are usually fine for prototyping and unreliable for production. The proof is in the SLO.
• Will the catalog include closed models? Right now it's open weights. Whether NVIDIA can broker access to GPT-class or Claude-class models would be the real platform play.

The bottom line

For developers and small teams, this is the easiest way to test 230B-parameter open models without owning a GPU rack or paying for a specialist inference provider. For enterprises, it's a reason to revisit the "build vs. buy vs. broker" question on your AI stack - NVIDIA just added a third option to the menu.

You can get your API key right now at build.nvidia.com. It's free. Registration takes about five minutes.