Free LLM API, No Key: Models You Can curl Right Now

Most "free LLM API" lists are a graveyard of expired trials, waitlists, and endpoints that want a credit card "for verification." This one is different in a verifiable way: every model below was called with a bare curl — no API key, no wallet, no account — on the day this article was published. You can re-run the check yourself in ten seconds:

curl https://blockrun.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/gpt-oss-120b",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

No auth header. Returns a standard OpenAI-shaped chat completion.

The free models

These are open-source models hosted free on NVIDIA's infrastructure and routed through BlockRun's gateway. Same endpoint, same request shape as any paid model — you just skip the payment header entirely.

Context windows and pricing (input: 0, output: 0) are queryable from the same API, so this table is checkable, not vibes: curl https://blockrun.ai/api/v1/models — also free, also keyless.

Quickstart in three languages

The endpoint is OpenAI-compatible, so every OpenAI client library works by changing one URL.

Python (official openai package):

from openai import OpenAI

client = OpenAI(
    base_url="https://blockrun.ai/api/v1",
    api_key="not-needed",  # the SDK requires a string; the API doesn't check it
)

r = client.chat.completions.create(
    model="nvidia/deepseek-v4-flash",
    messages=[{"role": "user", "content": "Summarize x402 in one sentence."}],
)
print(r.choices[0].message.content)

JavaScript / TypeScript:

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://blockrun.ai/api/v1",
  apiKey: "not-needed",
});

const r = await client.chat.completions.create({
  model: "nvidia/gpt-oss-120b",
  messages: [{ role: "user", content: "Hello" }],
});
console.log(r.choices[0].message.content);

Streaming works the standard way — add "stream": true to the body. Tool use and vision input (on nemotron-nano-12b-v2-vl) follow the OpenAI schema too.

The honest limits

Free means free, not infinite. Three things to know before you build on it:

Per-IP rate limits apply. Enough for development, prototypes, and light production; not enough to run a scraping farm. If you hit the limit, you'll get a clean 429.

Cold starts happen. Free capacity runs on shared NVIDIA infrastructure. The first call after a quiet period occasionally returns a FREE_MODEL_FAILED error — retrying once almost always succeeds. (We hit this ourselves while verifying this article; every model answered on retry.) Build in one retry and you'll rarely see it.

No SLA. The free lineup is re-validated continuously and models are swapped when upstream capacity ends. Treat the model list endpoint as the source of truth rather than hardcoding one model forever.

When you outgrow free

The upgrade path deliberately has no cliff: no account creation, no API key, no sales call. Paid models on the same endpoint are priced per call (from a $0.002 minimum per request) and settle in USDC over the x402 protocol — your agent signs a payment header instead of you signing up. Per-token rates on paid chat models match OpenRouter with no platform margin; the only add-on is a flat $0.001 transaction fee. The full lineup — GPT, Claude, Gemini, Grok, DeepSeek and more — is on the models page, with pricing public.

FAQ

Is there really no signup? Really. The curl at the top of this page is the entire onboarding. The API only checks for a payment header when you call a paid model.

Why is it free? The models are open-source and NVIDIA hosts them at no charge; BlockRun routes to them so agents get one endpoint for both free and paid calls. Free traffic is how most people first test the gateway — that's the business logic, and it doesn't require your email address to work.

Can I use these in production? For light workloads, yes — with the retry caveat above and the rate limits in mind. For anything revenue-bearing, the paid tier exists precisely because it comes with capacity that isn't best-effort.

What about privacy? Requests are routed to NVIDIA-hosted open-source models. Don't send secrets to any free LLM endpoint, ours included.