Free LLM API.
No key. No wallet.
9 open-source models, free through BlockRun. No key. No wallet. Six ways to call any of them.
curl https://blockrun.ai/api/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia/nemotron-3.5-lightning",
"messages": [{"role": "user", "content": "Hello"}]
}'5 dedicated pages.
One free endpoint.
Free Llama API.
NVIDIA retired every free Llama in November 2026. Meta's successor line is still free here: Muse Glimmer 30B, 131K context, a reasoning model. No key, no wallet, no subscription. Just call it.
Free Reasoning API.
Nemotron 3.5 Lightning (131K context, thinking mode) and Nemotron 3 Ultra 550B — 550B total / 55B active, the largest free model in the catalog. No key, no signup.
Free Mistral API.
Our free Mistral upstream is down — calls to nvidia/mistral-* still return 200, answered by a healthy free model. Meanwhile, two free open-weights coding models with no key and the same OpenAI-compatible endpoint.
Free Nemotron API. With vision.
Nemotron 3 Nano Omni — 31B / 3.2B active MoE, 256K context. The only free model that takes images, video, and audio. NVIDIA-reported: ChartQA 90.3, DocVQA 95.6, MMMU 70.8.
Free GPT-OSS API.
GPT-OSS, OpenAI's open-weights chat model family. The 20B variant, 128K context, hosted free on NVIDIA and called through BlockRun. (NVIDIA retired its free 120B hosting on 2026-09-03; calls to it now redirect.)
Free models,
free-tier terms.
Free models are served on free tiers whose terms may let the host use prompts to improve its service, so don't send sensitive data to a free model. We never sell your data.
- Free models run on free tiers whose terms may let the host use prompts to improve its service. Don't send sensitive data to a free model.
- Wallet in, prompt out. Pseudonymous by default — no email, no phone number, no identity documents.
- Read the code, audit the wire format, run it yourself. @blockrun/llm and blockrun-llm on npm and PyPI.
A free model can be busy.
Your call still returns 200.
Free capacity is borrowed, not reserved, so at any moment some of these models are saturated. When the one you named is, we answer with another free model instead of failing the request. On 2026-09-21 five of the six were busy and all six calls returned 200.
The reply always tells you which one answered. Read the model field rather than assuming you got the ID you sent:
# you asked for one model "model": "nvidia/nemotron-3.5-lightning" # the reply names what actually served it "model": "nvidia/nemotron-3-super-120b"
That is the trade for paying nothing. If you need a specific model to be the one that answers every time, that is what the pay-per-call side is for — same endpoint, same request, and the model you name is the model you get.
Frontier models are
pay-per-call. Same endpoint.
Claude Opus, GPT-5.6, Gemini, Grok, Kimi — 87 paid models on the same API. No subscription. No monthly minimum. Top up $5 in USDC and call anything.