NVIDIA NIM Framework Integrations

All integrations share two required changes — and only two:

Everything else — authentication header format, request/response schema, parsing logic — stays identical to how you already call OpenAI. [NVIDIA NIM Free Models Guide]

The free model highlighted across all examples is GLM-4, a bilingual (Chinese/English) model from Zhipu AI that handles reasoning, instruction-following, and code tasks well. [NVIDIA NIM Free Models Guide]

Claude Code reads credentials from environment variables. Set these in your shell or profile file:

export NVIDIA_NIM_API_KEY="nvapi-xxxxxxxxxxxxxxxx"
export NVIDIA_NIM_BASE_URL="https://integrate.api.nvidia.com/v1"

For custom scripts that Claude Code invokes inside an agentic loop, use the OpenAI Python SDK pointed at NIM:

from openai import OpenAI

client = OpenAI(
    api_key="nvapi-xxxxxxxxxxxxxxxx",
    base_url="https://integrate.api.nvidia.com/v1"
)

response = client.chat.completions.create(
    model="zhipuai/glm-4",
    messages=[{"role": "user", "content": "Explain this function in one paragraph."}],
    temperature=0.3,
    max_tokens=512
)

print(response.choices[0].message.content)

Because the response structure is identical to OpenAI's, existing parsing logic requires no changes. [NVIDIA NIM Free Models Guide]

LangChain accepts a custom base_url directly on ChatOpenAI:

from langchain_openai import ChatOpenAI

nim_llm = ChatOpenAI(
    model="zhipuai/glm-4",
    api_key="nvapi-xxxxxxxxxxxxxxxx",
    base_url="https://integrate.api.nvidia.com/v1",
    temperature=0.2
)

response = nim_llm.invoke("Summarize this document in three bullet points.")

In a LangGraph multi-agent graph, different nodes can be assigned different models, enabling a model tiering strategy where only high-stakes nodes use premium models:

from langgraph.graph import StateGraph
from langchain_openai import ChatOpenAI

# Expensive model for planning
planner = ChatOpenAI(model="gpt-4o", api_key="your-openai-key")

# Free NIM model for data extraction
extractor = ChatOpenAI(
    model="zhipuai/glm-4",
    api_key="nvapi-xxxxxxxxxxxxxxxx",
    base_url="https://integrate.api.nvidia.com/v1"
)

def planning_node(state):
    return {"plan": planner.invoke(state["task"])}

def extraction_node(state):
    return {"extracted": extractor.invoke(state["document"])}

[NVIDIA NIM Free Models Guide]

CrewAI requires the model name to be prefixed with openai/ when using an OpenAI-compatible third-party endpoint:

from crewai import LLM

nim_model = LLM(
    model="openai/zhipuai/glm-4",   # note the openai/ prefix
    api_key="nvapi-xxxxxxxxxxxxxxxx",
    base_url="https://integrate.api.nvidia.com/v1"
)

Assign this LLM to crew members handling lower-complexity tasks such as data normalisation, formatting, or boilerplate generation. [NVIDIA NIM Free Models Guide]

Pass a standard config_list with api_type set to "openai":

config_list = [
    {
        "model": "zhipuai/glm-4",
        "api_key": "nvapi-xxxxxxxxxxxxxxxx",
        "base_url": "https://integrate.api.nvidia.com/v1",
        "api_type": "openai"
    }
]

Supply this config to any AutoGen agent that does not require the most capable model in your setup. [NVIDIA NIM Free Models Guide]

For tools without SDK support, raw HTTP works against the same endpoint:

curl https://integrate.api.nvidia.com/v1/chat/completions \
  -H "Authorization: Bearer nvapi-xxxxxxxxxxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zhipuai/glm-4",
    "messages": [{"role": "user", "content": "Hello, what can you do?"}],
    "max_tokens": 256
  }'

[NVIDIA NIM Free Models Guide]

MindStudio offers a different path for teams who do not want to manage API keys, base URLs, or SDK wiring. It ships with 200+ models available out of the box, and model tiering — routing tasks to cheaper or free models automatically — is built into the visual workflow structure. It also includes 1,000+ pre-built integrations with tools such as Slack, HubSpot, Notion, and Google Workspace. [NVIDIA NIM Free Models Guide]

The diagram below shows how each framework sits between your application and the NVIDIA NIM hosted endpoint:

The model tiering strategy described in the source recommends routing tasks by complexity: [NVIDIA NIM Free Models Guide]

Applying this split can reduce API spend by 40–70% on typical agentic workflows without meaningful quality degradation on routed tasks. [NVIDIA NIM Free Models Guide]

• Every framework covered requires only two changes: the base URL (https://integrate.api.nvidia.com/v1) and the model ID. [NVIDIA NIM Free Models Guide]
• CrewAI is the only framework with a non-obvious quirk: the openai/ prefix on the model name. [NVIDIA NIM Free Models Guide]
• GLM-4 is a notable free-tier option, particularly strong for bilingual Chinese/English tasks. [NVIDIA NIM Free Models Guide]
• Free models have rate limits — add retry logic before running batch workloads. [NVIDIA NIM Free Models Guide]
• MindStudio removes all configuration overhead for teams building multi-model workflows without code. [NVIDIA NIM Free Models Guide]

For a broader introduction to the platform itself, see the NVIDIA NIM Free Models Guide. For context on the free-vs-paid decision, see Free vs. Paid AI Model Access: NVIDIA NIM vs. $20/month Subscriptions.