价格核实于 2026-07 · 每季度更新

云 GPU 与大模型 API

在家跑不动的模型，有三种运行方式：按小时租用 GPU、用免费 API 额度做原型，或按 token 付费。从最适合你工作负载的方式开始。

云端 AI 目录

由 Jakub Rusinowski 整理 · 条目核实于 2026-07-08

本站关注在自己的硬件上运行 AI——但云端出人意料地常常是正确的工具：用免费 API 做原型，租 GPU 做微调或先试后买，想要 Llama 级别的输出又不想买显卡时用托管开源模型。这是一份诚实的全景图，以及本地何时更划算的计算。

价格说明刻意保持粗略（额度经常变化），请以提供商页面为准。大多数是普通链接；RunPod 和 Vast.ai 是联盟链接。

免费大模型 API

真正免费的 API——有速率限制，但真实可用。足够在购买硬件或 token 之前做原型、学习和小规模工具。

NVIDIA NIM (build.nvidia.com)免费: Free hosted endpoints for Llama, Nemotron, Mistral & more with generous dev credits — sign-up only, no cardFree API credits for development; self-host NIM containers with NVIDIA AI Enterprise in productionTrying big open models (Llama 3.3 70B, Nemotron) over an OpenAI-compatible API before buying hardwareGoogle AI Studio免费: Gemini Flash free tier — the most generous standing free allowance of the big threeFree-of-charge tier with per-minute/per-day rate limits; paid tier unlocks higher limitsPrototyping against a frontier-family model at zero costGroq免费: Free developer tier with daily rate limits on open models (Llama, Qwen, Gemma)Pay-per-token on LPU hardware; famous for extreme tokens/secLatency-critical demos — hundreds of tokens/sec on 70B-class open modelsOpenRouter免费: Rotating set of :free model variants with daily request capsOne API key, 300+ models, pass-through pricing per modelComparing many models through one OpenAI-compatible endpoint before committingMistral La Plateforme免费: Free experiment tier with rate limitsPer-token pricing on Mistral models; EU-hostedEU-hosted API access with a free on-ramp — relevant for GDPR-sensitive prototypingCerebras Inference免费: Free developer tier with daily token capsPer-token pricing on wafer-scale hardware; very high tokens/secFastest hosted open-model inference alongside Groq — good for agent loopsGitHub Models免费: Free playground + API rate limits for GPT-4o, Llama, Phi, Mistral with any GitHub accountIncluded with a GitHub account; production use routes to Azure AIZero-signup-friction experimentation if you already live on GitHubCloudflare Workers AI免费: Daily free allocation of neurons on every Cloudflare accountPer-neuron (unit) pricing beyond the free allocation; runs on Cloudflare edge GPUsSprinkling small-model inference (Llama 8B class) into edge apps without managing serversHugging Face Inference Providers免费: Small monthly free credit allowance on every HF accountRoutes to partner providers at pass-through prices; monthly included credits on free & PRO plansCalling almost any open-weight model on the Hub without picking a vendor first

云 GPU 租用

按小时租 GPU 而不是购买。适合微调、批处理任务，以及“在花 2000 美元装机前先试试 70B 模型”。

RunPodRTX 4090 ≈ $0.3–0.7/hr, A100/H100 by the hour; serverless per-second billing availableRenting a 24–80 GB GPU for hours-long jobs: fine-tuning, batch inference, trying 70B before buying联盟链接Vast.aiMarketplace of hosted GPUs — often the cheapest per hour (interruptible options)Cheapest possible GPU-hours when you can tolerate marketplace variability联盟链接LambdaOn-demand A100/H100/B200 instances and clusters, per-hour billingSerious training/fine-tuning runs on datacenter GPUs with clean imagesPaperspace (DigitalOcean)免费: Free-tier Gradient notebooks with entry GPUs, subject to availabilityPer-hour GPU machines + Gradient notebooks; some free-GPU notebook availabilityNotebook-first experimentation with an easy upgrade pathHyperstackAggressively priced A100/H100 per-hour instances in EU/NA regionsEU-region datacenter GPUs at low hourly ratesTensorDockMarketplace pricing; consumer + datacenter GPUs from ~$0.2/hrBudget hourly GPUs with per-minute billingGoogle Colab免费: Free T4 GPU notebook sessions — the classic zero-cost on-rampFree notebooks with T4 GPUs (availability-based); Pro tiers add better GPUs and runtimesLearning, small fine-tunes (LoRA on 7–8B), and quick experiments without any setup

托管开源模型 API

由专业厂商按 token 提供开源权重模型（Llama、Qwen、Mistral、DeepSeek）——介于前沿 API 和自有硬件之间。

Together AI免费: Small starter credit for new accountsPer-token pricing on 100+ open models (Llama 3.1 8B ≈ $0.18/M tokens)Production open-model serving with serious throughput and fine-tuning endpointsFireworks AI免费: Starter credit for new accountsPer-token open-model serving; fast function-calling-optimized deploymentsLow-latency production inference on open models, incl. dedicated deploymentsDeepInfraAmong the lowest per-token prices for open models; OpenAI-compatibleCheapest hosted per-token rates on Llama/Qwen/Mistral-class modelsReplicatePer-second (hardware-based) or per-token pricing; huge community model catalogRunning community models (image, audio, video, LLM) without packaging them yourselfNovita AI免费: Small trial creditBudget per-token open-model APIs + GPU instancesLow-cost open-model APIs with an Asia-Pacific footprint

前沿模型 API

闭源前沿模型。质量最高、零配置、按 token 计费——所有本地方案都会拿它作对比。

OpenAI APIGPT-4o $2.50/M input, $10/M output; GPT-4o Mini $0.15/$0.60Frontier quality and the broadest tooling ecosystemAnthropic APIClaude Sonnet $3/M input, $15/M output; Haiku $0.80/$4Best-in-class coding and long-context analysisGoogle Gemini API免费: Free tier via Google AI Studio (see above)Gemini Flash $0.10/M input, $0.40/M output; Pro $1.25/$5Huge context windows and the lowest frontier-family prices

作为亚马逊联盟成员，我们从符合条件的购买中获得收入。云 GPU 链接为推荐链接——我们可能获得佣金，您无需额外付费。

本地还是云端？

附带统一回本计算器的对比文章——如实说明各自的优势。云端价格核实于 2026-03-01。

从这里开始Local AI vs Cloud API Costs: The Real Break-Even Math (2026)Local AI vs cloud API costs compared with real numbers: hardware, electricity, and per-token prices. Break-even calculator shows when a GPU pays for itself.

常见问题

哪里可以免费通过 API 使用 AI 模型？

2026 年最有用的长期免费额度：NVIDIA NIM（build.nvidia.com）为 Llama、Nemotron 和 Mistral 级模型提供免费托管端点；Google AI Studio 提供慷慨的 Gemini Flash 免费额度；Groq 和 Cerebras 为开发者提供有速率限制、速度极快的开源模型免费额度；OpenRouter 提供轮换的免费模型；GitHub Models 任何 GitHub 账户都能使用。

租用 AI GPU 最便宜的方式是什么？

交易市场最便宜：Vast.ai 和 TensorDock 提供消费级 GPU（RTX 3090/4090 级别），约每小时 0.20–0.40 美元；RunPod 提供按秒计费的无服务器方案；Lambda/Hyperstack 覆盖数据中心 A100/H100。免费试验方面，Google Colab 仍提供免费的 T4 笔记本会话。

应该租云 GPU 还是自己买？

偶尔使用（微调、先试 70B 模型再决定）就租——每次实验几美元。持续使用就买：每天推理几个小时，1000–2600 美元的配置通常在几个月内就比租用便宜。我们的本地与云端对比页面附带回本计算器。

托管开源模型 API 和前沿模型 API 有什么区别？

托管开源模型的提供商（Together、Fireworks、DeepInfra、Groq）按 token 提供 Llama、Qwen 等开源权重模型——通常比前沿 API 便宜 10–50 倍，而且以后可以把同一个模型迁移到自己的硬件上。前沿 API（OpenAI、Anthropic、Google）提供质量上限最高的闭源模型。

By budget

By use case

By team size

By privacy & compliance

更想拥有自己的硬件？

工具