NVIDIA's Nemotron 3 Super 120B (nvidia/nemotron-3-super-120b-a12b) is one of the most capable open-weight models available today — and it runs at $0 input / $0 output on NVIDIA NIM through InferAll, within the free-plan daily request caps (50 chat / 100 text / day, reset 00:00 UTC). Activate via the $5 starter pack at /billing — the $5 becomes spendable balance for premium providers (OpenAI, Anthropic, Google at the published per-token rate with zero markup).
Here's how to call it using the standard OpenAI SDK:
from openai import OpenAI
client = OpenAI(
base_url="https://api.inferall.ai/v1",
api_key="ifu_your_key_here", # get one at inferall.ai/keys
)
response = client.chat.completions.create(
model="nvidia/nemotron-3-super-120b-a12b",
messages=[{"role": "user", "content": "What makes a good system prompt?"}],
max_tokens=512,
)
print(response.choices[0].message.content)
That's it. The same call works with any OpenAI-compatible library — LangChain, LlamaIndex, LiteLLM, CrewAI, and any other framework that accepts an openai_api_base override.
Why Nemotron 120B
Scale. At 120B parameters, Nemotron outperforms many models twice its cost on reasoning, instruction-following, and long-context tasks. It scores well on coding and math benchmarks without the rate-limit friction of commercial APIs.
Free on NIM. NVIDIA hosts Nemotron on their DGX Cloud infrastructure via NIM (NVIDIA Inference Microservices), which InferAll exposes at $0. There's no inference cost for us to pass through, so it stays free within the allowance.
OpenAI-compatible. The model is served through InferAll's OpenAI-compatible endpoint — you get standard ChatCompletion responses, streaming, tool use, and JSON mode, all working with whatever OpenAI client you already have.
TypeScript / Node.js
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.inferall.ai/v1",
apiKey: process.env.INFERALL_API_KEY,
});
const response = await client.chat.completions.create({
model: "nvidia/nemotron-3-super-120b-a12b",
messages: [{ role: "user", content: "Explain backpropagation." }],
});
Streaming
with client.chat.completions.create(
model="nvidia/nemotron-3-super-120b-a12b",
messages=[{"role": "user", "content": "Write a poem about distributed systems."}],
stream=True,
) as stream:
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")
Claude Code / Cline / Cursor
Set these two environment variables and any Anthropic-compatible agent routes through InferAll:
export ANTHROPIC_BASE_URL=https://api.inferall.ai
export ANTHROPIC_API_KEY=ifu_your_key_here
When your agent tries to use claude-opus-4-8, the gateway maps it to Nemotron (opus-class model, same relative capability tier). Works immediately, no configuration changes.
Comparing the free models
All of these are $0 on InferAll, hosted on NVIDIA NIM:
| Model | Size | Best for |
|---|---|---|
nvidia/nemotron-3-super-120b-a12b |
120B | Complex reasoning, coding, long context |
meta/llama-3.1-70b-instruct |
70B | General chat, instruction following |
meta/llama-3.1-8b-instruct |
8B | Fast responses, simple tasks |
mistralai/mistral-medium-3.5-128b |
46.7B (MoE) | Speed + quality balance |
See the live model list for all 40+ free models.
Get a key
inferall.ai/keys — sign up free and start calling (25 free NIM calls before any payment; the $5 starter pack unlocks ongoing use) at /billing. The $5 becomes spendable balance: 40+ open NIM models stay $0 in/out against it (within the free-plan daily request caps); premium providers (OpenAI, Anthropic, Google) bill at the provider's published per-token rate with zero markup.
# Verify the model is live
curl https://api.inferall.ai/ai/v1/models | jq '."nvidia/nemotron-3-super-120b-a12b"'