NVIDIA NIM 通过 NVIDIA 云基础设施提供优化的 AI 模型推理 API，支持大语言模型、视觉和向量模型。
integrate.api.nvidia.com完成站点认证，解锁以下权益：
排名基于社区提交的测试数据与定期健康探测，仅供参考，非官方数据。
按 59 个模型行对比最新 audit、最新测速、吞吐、延迟与按 token 计费价格。
| 模型 | 输入 ($/M) | 输出 ($/M) | 检测 | 速度 | 延迟 |
|---|---|---|---|---|---|
| - | - | 10010086100 | 81.1 t/s | 1.05 s | |
google/diffusiongemma-26b-a4b-it | - | - | — | 697.7 t/s | 0.70 s |
nvidia/nemotron-3-ultra-550b-a55b | - | - | — | 79.4 t/s | 1.70 s |
| - | - | — | 118.6 t/s | 15.57 s | |
| - | - | — | 46.8 t/s | 10.40 s | |
nvidia/nemotron-3-nano-omni-30b-a3b-reasoning | - | - | — | 274.5 t/s | 0.66 s |
google/gemma-4-31b-it | - | - | 1006886100 | 20.4 t/s | 26.82 s |
meta/llama-3.1-8b-instruct | - | - | — | 145.7 t/s | 0.21 s |
meta/llama-3.3-70b-instruct | - | - | — | 28.1 t/s | 5.27 s |
mistralai/mistral-small-4-119b-2603 | - | - | — | 120.9 t/s | 0.29 s |
nvidia/nemotron-mini-4b-instruct | - | - | — | 86.6 t/s | 0.45 s |
nvidia/nemotron-3-nano-30b-a3b | - | - | — | 207.0 t/s | 1.13 s |
| - | - | — | 83.2 t/s | 0.39 s | |
| - | - | 727280100 | 21.4 t/s | 1.34 s | |
| - | - | 1006878100 | 22.8 t/s | 14.36 s | |
| - | - | — | 42.3 t/s | 32.01 s | |
| - | - | 848480100 | 82.4 t/s | 9.31 s | |
nvidia/nemotron-3-super-120b-a12b | - | - | — | 28.5 t/s | 31.58 s |
marin/marin-8b-instruct | - | - | — | 84.2 t/s | 0.44 s |
qwen/qwen3-next-80b-a3b-thinking | - | - | — | 106.6 t/s | 9.02 s |
meta/llama-4-maverick-17b-128e-instruct | - | - | — | 92.8 t/s | 0.27 s |
institute-of-science-tokyo/llama-3.1-swallow-70b-instruct-v0.1 | - | - | — | 19.0 t/s | 0.49 s |
meta/llama-3.2-90b-vision-instruct | - | - | — | 16.9 t/s | 0.69 s |
| - | - | — | 18.5 t/s | 0.50 s | |
meta/llama-3.1-405b-instruct | - | - | — | 22.7 t/s | 5.87 s |
qwen/qwen3-coder-480b-a35b-instruct | - | - | — | 49.3 t/s | 2.07 s |
meta/llama3-70b-instruct | - | - | — | 37.5 t/s | 0.47 s |
| - | - | — | 31.3 t/s | 16.79 s | |
z-ai/glm5 | - | - | — | 24.9 t/s | 35.20 s |
stockmark/stockmark-2-100b-instruct | - | - | — | 55.3 t/s | 0.74 s |
nvidia/llama-3.1-nemotron-ultra-253b-v1 | - | - | — | 45.2 t/s | 0.21 s |
| - | - | — | 45.2 t/s | 2.90 s | |
| - | - | 848480100 | 28.0 t/s | 13.61 s | |
| - | - | — | 78.8 t/s | 3.19 s | |
| - | - | — | 81.8 t/s | 4.79 s | |
| - | - | — | 47.9 t/s | 18.53 s | |
z-ai/glm4.7 | - | - | — | 57.0 t/s | 27.72 s |
| - | - | — | 86.4 t/s | 2.88 s | |
01-ai/yi-large | - | - | — | 43.7 t/s | 0.22 s |
| - | - | — | 55.6 t/s | 0.29 s | |
| - | - | — | 41.4 t/s | 2.36 s | |
google/gemma-2-27b-it | - | - | — | 43.7 t/s | 0.23 s |
google/gemma-3-27b-it | - | - | — | 50.8 t/s | 0.47 s |
meta/llama-3.1-70b-instruct | - | - | — | 51.2 t/s | 0.23 s |
microsoft/phi-3-medium-128k-instruct | - | - | — | 18.1 t/s | 0.49 s |
microsoft/phi-4-mini-flash-reasoning | - | - | — | 74.2 t/s | 0.46 s |
mistralai/mistral-small-24b-instruct | - | - | — | 29.7 t/s | 0.49 s |
mistralai/mixtral-8x22b-instruct-v0.1 | - | - | — | 89.7 t/s | 0.22 s |
| - | - | — | 45.1 t/s | 0.58 s | |
nvidia/llama-3.1-nemotron-70b-instruct | - | - | — | 52.2 t/s | 0.23 s |
nvidia/llama-3.3-nemotron-super-49b-v1.5 | - | - | — | 57.5 t/s | 11.11 s |
| - | - | — | 17.9 t/s | 4.49 s | |
| - | - | — | 41.3 t/s | 19.45 s | |
| - | - | — | 152.8 t/s | 0.95 s | |
| - | - | — | 162.7 t/s | 1.12 s | |
| - | - | — | 44.5 t/s | 21.25 s | |
| - | - | — | 87.6 t/s | 8.96 s | |
| - | - | — | 41.0 t/s | 0.49 s | |
| - | - | — | 34.0 t/s | 0.59 s |
当前显示 59 / 59 个模型行
| 时间 | 模型 | 速度 | 延迟 |
|---|---|---|---|
| Jul 5, 05:47 AM | z-ai/glm-5.2 | 24.30 tok/s | 1.40s |
| Jul 3, 03:37 AM | z-ai/glm-5.2 | 137.86 tok/s | 0.70s |
| Jul 2, 04:51 AM | deepseek-ai/deepseek-v4-flash | 13.40 tok/s | 57.53s |
| Jul 2, 04:49 AM | qwen/qwen3.5-122b-a10b | 58.72 tok/s | 1.67s |
| Jul 2, 04:42 AM | deepseek-ai/deepseek-v4-pro | 21.19 tok/s | 1.40s |
| Jul 2, 04:41 AM | mistralai/mistral-small-4-119b-2603 | 120.92 tok/s | 0.29s |
| Jul 2, 04:37 AM | mistralai/mistral-medium-3.5-128b | 46.84 tok/s | 10.40s |
| Jul 2, 04:34 AM | nvidia/nemotron-3-ultra-550b-a55b | 36.46 tok/s | 2.98s |
| Jul 2, 04:33 AM | moonshotai/kimi-k2.6 | 68.46 tok/s | 10.97s |
| Jul 2, 04:32 AM | google/diffusiongemma-26b-a4b-it | 697.68 tok/s | 0.70s |
用 330 个 LMSpeed 信号，把 NVIDIA NIM 的 6 个相近 API 替代服务商放在一起比较：共享模型覆盖、价格、实测速度、可用性和免费模型。
| 服务商 | 对比理由 | 模型数 | 免费项 | 均价 | 速度 | 30 天可用性 |
|---|---|---|---|---|---|---|
| NVIDIA NIM nvidia-nim NVIDIA NIM provides optimized AI model inference APIs for LLMs, vision, and embedding models through NVIDIA cloud infrastructure. | 当前服务商基线 | 53 | 0 | N/A | 66 tok/s | 98.9% |
| OpenRouter openrouter A unified API interface providing access to over 300 models from 60+ providers, including OpenAI, Anthropic, and Google. |
| 221 | 0 | N/A | 86 tok/s | 99.1% |
api-kriora-com Provides OpenAI-compatible APIs and managed GPU instances for deploying and scaling open-source AI models. |
| 9 | 0 | N/A | 565 tok/s | 99.1% |
api-fireworks-ai Fireworks AI provides a cloud platform for running and fine-tuning open-source AI models with optimized inference for production applications. |
| 5 | 0 | N/A | 139 tok/s | 99.1% |
nebius-ai-studio Nebius AI Studio provides managed inference APIs for open-weight and proprietary AI models, with developer-focused tooling for building and scaling AI applications. |
| 1 | 0 | N/A | 74 tok/s | 99.1% |
siliconflow Provides cost-effective generative AI cloud services based on open-source models for text, image, video, and audio generation. |
| 60 | 0 | N/A | 53 tok/s | 76.8% |
api-model-yungnet-cn Yungnet Model Open Platform is a China-based unified LLM API gateway with an OpenAI-compatible relay at api.model.yungnet.cn, focusing on flagship Chinese models. |
| 16 | 0 | N/A | 49 tok/s | 99.1% |
