⚠️ 待更新·2026-08-29核验 · 更新时间待核验 · 本文信息可能已过期，请以官方文档为准
更新时间：2026-08-29 · 核验状态：待更新 · 官方溯源待补
引言
NVIDIA NIM 把主流开源模型打包成标准化推理容器并提供云端托管端点。注册 build.nvidia.com 可获 1000 credits 免费额度，覆盖数十个模型。
可用模型与端点
端点：https://integrate.api.nvidia.com/v1/chat/completions（OpenAI 兼容）。
meta/llama-3.1-405b-instructmeta/llama-3.1-70b-instructnvidia/llama-3.1-nemotron-70b-instructmistralai/mistral-nemo-12b-instructqwen/qwen2.5-7b-instructdeepseek-ai/deepseek-r1
调用示例
from openai import OpenAI
client = OpenAI(
base_url="https://integrate.api.nvidia.com/v1", api_key="nvapi-...")
resp = client.chat.completions.create(
model="deepseek-ai/deepseek-r1",
messages=[{"role": "user", "content": "用 Python 实现快速排序"}],
temperature=0.6, max_tokens=1024)
print(resp.choices[0].message.content)
注意事项
NIM 与英伟达硬件深度优化，延迟低、吞吐高。1000 credits 约百万 token，适合评测与原型。credits 用尽返回 402；405B 消耗快；敏感数据走本地 NIM 容器（nvcr.io/nim/...）。
⚠️ Pending Update · 2026-08-29 Verification · Content may be outdated, please refer to official docs
Updated: 2026-08-29 · Status: Pending Verification
Introduction
NVIDIA NIM (NVIDIA Inference Microservices) packages popular open-source models into standardized inference containers and hosts them in the cloud. Signing up at build.nvidia.com grants 1000 credits, which cover dozens of models and translate to roughly a million tokens of inference — enough for thorough evaluation and prototyping before you commit to a paid plan. Because NIM is co-designed with NVIDIA hardware, it consistently delivers lower latency and higher throughput than generic inference clouds, which makes it attractive even when the free credits run out. NIM also shines for consistency between cloud and self-hosted deployments: the same container image runs locally and in the cloud, so you can prototype on the free cloud endpoint and migrate to your own GPUs later without changing a line of application code. That portability is rare among inference providers and is a meaningful advantage for teams that expect to bring workloads on-prem eventually.
Available Models and Endpoint
Endpoint: https://integrate.api.nvidia.com/v1/chat/completions (OpenAI-compatible, so any OpenAI SDK works out of the box). Common free-tier models include:
meta/llama-3.1-405b-instruct — frontier-scale general model.meta/llama-3.1-70b-instruct — balanced quality and cost.nvidia/llama-3.1-nemotron-70b-instruct — NVIDIA-tuned for helpfulness.mistralai/mistral-nemo-12b-instruct — compact multilingual model.qwen/qwen2.5-7b-instruct — strong bilingual chat.deepseek-ai/deepseek-r1 — reasoning chain, strong at math and code.
Call Example
from openai import OpenAI
client = OpenAI(
base_url="https://integrate.api.nvidia.com/v1", api_key="nvapi-...")
resp = client.chat.completions.create(
model="deepseek-ai/deepseek-r1",
messages=[{"role": "user", "content": "Implement quicksort in Python."}],
temperature=0.6, max_tokens=1024)
print(resp.choices[0].message.content)
Caveats
The 1000-credit bonus is generous for evaluation but burns quickly on the 405B model — a single long call can consume dozens of credits, so prefer 70B or smaller for iterative testing. When credits run out, the API returns HTTP 402 rather than a graceful error, so build credit-balance checks into your client to fail soft. Although NIM does not use request data for training, regulated workloads should still run on a self-hosted NIM container (nvcr.io/nim/...), which exposes the identical OpenAI-compatible API on your own GPU and keeps data entirely within your network. Finally, NIM exposes non-chat tasks too — embeddings, reranking, and vision models share the same /v1 endpoint family, which means a single API key and client can serve a full RAG pipeline. The free credits apply across all task types, so you can benchmark embeddings quality and chat quality in the same evaluation run without juggling multiple accounts. Budget your evaluation runs against the credit balance from day one, because the free tier is most useful when spent deliberately on benchmarking rather than frittered on exploratory calls.