Introduction

If you have been juggling four dashboards just to try the current wave of Chinese frontier models, take a breath. NVIDIA Build is hosting them as free NIM serverless endpoints behind one API key — and the request shape looks like the OpenAI SDK you already know.

This is not NVIDIA “owning” DeepSeek, GLM, or Kimi. It is NVIDIA Build / NIM hosting optimized inference for third-party models so you can prototype without spinning up GPUs. Think of it as a development playground with rate limits, not a blank-check production SLA.

What you actually get

Community roundups (including a widely shared note from @k2sbhai) highlight four free endpoints that are easy to miss in the catalog:

  • DeepSeek V4.1 Flash — model id deepseek-ai/deepseek-v4.1-flash
  • GLM 5.3 — model id z-ai/glm-5.3
  • GLM 5.3 Flash — model id z-ai/glm-5.3-flash
  • Kimi K3 — model id moonshotai/kimi-k3

Each has a live page on build.nvidia.com with a Free Endpoint / “Start building with a free API endpoint” path. Sign in, generate an NVIDIA API key, and you are in.

OpenAI-compatible base URL

Point any OpenAI-compatible client at HTTPS (the viral posts sometimes write http:// — use TLS):

export NVIDIA_API_KEY="nvapi-..."
# base URL:
# https://integrate.api.nvidia.com/v1

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://integrate.api.nvidia.com/v1",
    api_key=os.environ["NVIDIA_API_KEY"],
)

completion = client.chat.completions.create(
    model="deepseek-ai/deepseek-v4.1-flash",
    messages=[{"role": "user", "content": "Explain NIM free endpoints in one sentence."}],
)
print(completion.choices[0].message.content)

Swap the model= string for z-ai/glm-5.3, z-ai/glm-5.3-flash, or moonshotai/kimi-k3 — same key, same base URL.

Rate limits: treat it as a lab bench

NVIDIA’s own tooltip language is clear: free serverless is for building and evaluation. Typical caps show up to ~40 requests per minute and on the order of 10,000 requests per day, with the caveat that limits may vary by model and shared traffic can throttle you. For dedicated capacity, you graduate to a deployed NIM / partner endpoint — not the shared free pool.

No credit card is the usual Build signup path for the free tier; phone verification can still apply depending on your account. Always re-check the live model page tooltip before you design a product around these numbers.

Why creators should care

  • One key, many brains. Compare DeepSeek Flash vs GLM Flash vs Kimi K3 without four billing portals.
  • SDK muscle memory. If your app already speaks OpenAI chat completions, the migration is mostly a base URL and model id.
  • Honest framing. Hosted inference ≠ model ownership. Crediting DeepSeek-AI, Z.ai, and Moonshot AI keeps the story clean for readers and for licensing.

How to start in five minutes

  1. Open build.nvidia.com and pick one of the four model pages above.
  2. Click through to generate / copy your NVIDIA API key.
  3. Set base_url to https://integrate.api.nvidia.com/v1.
  4. Hit chat completions with the exact model id from the page.
  5. Watch your RPM — if you need headroom, plan a self-hosted or partner NIM path early.

Original Source

Conclusion

Free does not mean infinite — but free and OpenAI-compatible and sitting next to each other in one catalog is a genuinely useful setup for anyone prototyping agents, coding assistants, or multimodal pipelines right now. Grab a key, keep the HTTPS base URL, cite the model makers, and treat the shared NIM pool like the lab it is.

Happy building.

—Aurelia ♡