One free API key, four frontier Chinese models, zero card. Verified straight off each model’s own NVIDIA page — one of the five originally floating around this claim isn’t actually free, so it’s cut here.
Grab it:
- DeepSeek-V4-Flash-0731 — 304B MoE, 1M-token context, coding + agentic work
- MiniMax-M3 — vision-language, long-form video + long-horizon coding
- Qwen3.5-397B-A17B — 397B hybrid MoE, text/image/video, 201 languages
- Kimi-K3 — 2.8T MoE, 1M-token context, agentic coding + visual understanding
Set up (one key, all four):
- Account at build.nvidia.com — email only, no card
- Generate an NVAPI key from your account
- Base URL:
integrate.api.nvidia.com/v1— OpenAI-compatible, drop it into any existing tool - Pick the model name from the links above and call it
proof: each model page above states “governed by the NVIDIA API Trial Terms of Service” — that’s NVIDIA’s own free-trial tier, not a paid partner route.
⚠️ GLM-5.1 was in the original list — it's not actually on the free key
GLM-5.1’s own NVIDIA page shows Free Endpoint: Deprecated — it only offers Partner Endpoints or Self-Hosted Deployment now, both of which need your own paid setup elsewhere. It doesn’t share the free NVAPI key path the other four use, so it’s left off this list rather than sending you to a wall.
What NIM actually is, if you're new to this
NIM = NVIDIA’s hosted inference service — one endpoint, swap the model name to switch between 100+ models in their catalog, most on this same free trial tier. Longer background: Free NVIDIA AI Models topic.
Four keys become one. The fifth was never actually free — now you know before you build on it.
!