Verified 5 days ago · 2026-10-15
Hugging Face is an AI development platform offering model hosting, collaboration, inference APIs, and dedicated model deployment services.
Hugging Face official website ↗
For research and experimentation compare Free / PRO; for team collaboration compare Team / Enterprise; for a unified multi-model API use Inference Providers; for dedicated production deployment see Inference Endpoints.
Subscriptions
Subscription fee and monthly compute credits
- Free
- No fixed monthly allowance; buy credits under the current billing rules
- PRO
- USD 2 / month · compute credits, not tokens or chat counts
- Team
- USD 2 / user / month · compute credits, not tokens or chat counts
- Enterprise
- USD 2 / user / month · compute credits, not tokens or chat counts
Membership fees are not API balances; additional usage and dedicated deployments are billed separately.
Inference Providers
Billing, entitlements and official evidence
HF Routed Billing: Hugging Face bills the request, the selected provider runs inference, and the original model developer is identified separately.
Hugging Face routing markup: 0%
Hugging Face bills routed usage at provider rates with no added markup. Eligible included compute credits apply before paid usage; free accounts must purchase credits.
Free users must purchase prepaid credits before paid usage.
Custom provider API key
Provider Direct Billing: the provider bills directly; Hugging Face included credits do not apply. A custom provider key is billed directly by that provider. Hugging Face does not charge the call and its included credits do not apply.
| Model / developer | Inference provider / billing platform | Input | Output | Billing unit / modalities | Evidence |
|---|---|---|---|---|---|
GPT-OSS 120Bopenai/gpt-oss-120bOpenAI Existing model record | together Hugging Face | USD 0.15 | USD 0.6 | per million tokens text | Official source · 2026-10-15T08:00:53.716Z |
GPT-OSS 120Bopenai/gpt-oss-120bOpenAI Existing model record | novita Hugging Face | USD 0.05 | USD 0.25 | per million tokens text | Official source · 2026-10-15T08:00:55.891Z |
Llama 3.3 70B Instructmeta-llama/Llama-3.3-70B-InstructMeta Existing model record | together Hugging Face | USD 1.04 | USD 1.04 | per million tokens text | Official source · 2026-10-15T08:00:53.716Z |
No verified rates match these filters.
Compare the same currency, unit and routing conditions. Rates depend on the selected inference provider; credits do not change the model rate.
Supported tasks depend on the selected model and inference provider. Select the organization using bill_to or X-HF-Bill-To to use the organization's shared credits. Credits are issued monthly; no rollover or expiry guarantee was established from the checked pages. Availability is based on official live provider listings; authenticated inference was not tested.
Official source conflicts and open questions
- Hugging Face routed rates and Groq direct rates currently differ for GPT-OSS. Those Groq routes are omitted here pending clarification. /v1/models/openai/gpt-oss-120b · /v1/models/openai/gpt-oss-20b · /docs/models
- docs/inference-providers/index · 2026-10-15T08:05:30.644Z
- docs/inference-providers/pricing · 2026-10-15T08:05:30.644Z
- docs/inference-providers/hub-api · 2026-10-15T08:05:30.644Z
Inference Endpoints
Billing, entitlements and official evidence
| Instance | Hardware | System RAM / GPU or accelerator memory | Hourly rate | Billing and conditions | Evidence |
|---|---|---|---|---|---|
| AWS Intel Sapphire Rapids x1 aws | CPU · Intel Sapphire Rapids · 1 vCPU | 2 GB / — | USD 0.033 / hour / replica | Billed by the minute Billed by the minute. Region, capacity and quota must be confirmed in the deployment console. | Official source · 2026-10-15T07:57:38.509Z |
| AWS NVIDIA L4 x1 aws | GPU · NVIDIA L4 × 1 | — / 24 GB | USD 0.8 / hour / replica | Billed by the minute 24 GB is GPU memory; system RAM is not specified in this table. Billed by the minute; quota and capacity apply. | Official source · 2026-10-15T07:57:38.509Z |
| AWS Inferentia2 x1 aws | Accelerator · AWS Inferentia2 × 1 | 14.5 GB / 32 GB | USD 0.75 / hour / replica | Billed by the minute 32 GB accelerator memory and 14.5 GB system RAM. Compatible inference engine, capacity and quota required. | Official source · 2026-10-15T07:57:38.509Z |
Hourly rates are not token prices. Charges depend on billable time, running replicas and scaling; consult the documented instance lifecycle. No full-month usage or free scaling is assumed.
Active Hugging Face subscription and credit card required according to the current Endpoints pricing page. Instance capacity is quota-dependent; region availability must be checked in the deployment console. Compute for successfully deployed endpoints is charged during initialization and running. Replica scale-up time is billable. Scale-to-zero must be configured; it requires minimum replicas of zero and causes a cold start. Hourly rate × sum of billable replica minutes / 60. For autoscaling, include each additional replica's actual billable lifetime. Minute granularity is documented; partial-minute rounding and a universal minimum duration remain unconfirmed. Estimates may differ from invoices.
Official source conflicts and open questions
- The main pricing page rounds the CPU starting rate, and a monthly example uses an older rate. The prices here follow the current detailed instance table. /pricing · /docs/inference-endpoints/pricing
- docs/inference-endpoints/index · 2026-10-15T08:05:32.985Z
- docs/inference-endpoints/pricing · 2026-10-15T08:05:32.985Z
- docs/inference-endpoints/guides/autoscaling · 2026-10-15T08:05:32.985Z
个人研究与模型尝试可比较 Free / PRO；团队协作可比较 Team / Enterprise；多模型 API 选择 Inference Providers；生产独立部署查看 Inference Endpoints。
会员订阅
订阅费与每月计算额度
- Free
- 无固定每月赠送额度；调用前按官方规则购买 Credits
- PRO
- USD 2 / 月 · 计算额度，不是 Token 或聊天次数
- Team
- USD 2 / 用户 / 月 · 计算额度，不是 Token 或聊天次数
- Enterprise
- USD 2 / 用户 / 月 · 计算额度，不是 Token 或聊天次数
会员费不等于 API 余额；额外用量和专用部署单独计费。
模型推理 API
计费、权益与官方证据
HF Routed Billing：由 Hugging Face 结算，实际推理由选定服务商执行，模型开发商另行标识。
Hugging Face 路由加价: 0%
Hugging Face 统一路由并结算，按实际 Provider 单价收取，无额外加价；免费账户须购买额度。
Free 用户需要先购买预付 Credits。
自备 Provider API Key
Provider Direct Billing：第三方服务商直接结算，不使用 Hugging Face 赠送额度。 自带 Provider Key 时由上游直接计费，Hugging Face 赠送额度不适用。
| 模型 / 开发商 | 推理服务商 / 结算平台 | 输入 | 输出 | 计费单位 / 模态 | 证据 |
|---|---|---|---|---|---|
GPT-OSS 120Bopenai/gpt-oss-120bOpenAI 已有模型记录 | together Hugging Face | USD 0.15 | USD 0.6 | 每百万 Tokens text | 官方来源 · 2026-10-15T08:00:53.716Z |
GPT-OSS 120Bopenai/gpt-oss-120bOpenAI 已有模型记录 | novita Hugging Face | USD 0.05 | USD 0.25 | 每百万 Tokens text | 官方来源 · 2026-10-15T08:00:55.891Z |
Llama 3.3 70B Instructmeta-llama/Llama-3.3-70B-InstructMeta 已有模型记录 | together Hugging Face | USD 1.04 | USD 1.04 | 每百万 Tokens text | 官方来源 · 2026-10-15T08:00:53.716Z |
没有匹配的已核验报价。
仅比较同币种、同单位和相同路由条件。价格取决于所选推理服务商；Credits 抵扣不改变模型单价。
统一模型 API，支持文本、图像、视频、音频和嵌入等任务，具体取决于模型及推理服务商。计费平台、执行推理方和模型开发者分别列明。需要账户与访问令牌；调用测试未使用付费凭据。
官方资料冲突与待核验项
- Groq 的 HF 路由报价与 Groq 官网报价不同，暂不纳入已核验模型报价，保留待核实冲突。 /v1/models/openai/gpt-oss-120b · /v1/models/openai/gpt-oss-20b · /docs/models
- docs/inference-providers/index · 2026-10-15T08:05:30.644Z
- docs/inference-providers/pricing · 2026-10-15T08:05:30.644Z
- docs/inference-providers/hub-api · 2026-10-15T08:05:30.644Z
专用模型部署
计费、权益与官方证据
| 实例 | 硬件 | 系统内存 / 显存或加速器内存 | 小时单价 | 结算与条件 | 证据 |
|---|---|---|---|---|---|
| AWS Intel Sapphire Rapids x1 aws | CPU · Intel Sapphire Rapids · 1 vCPU | 2 GB / — | USD 0.033 / 小时 / 副本 | 按分钟结算 按分钟结算；区域、实例容量与配额以部署控制台为准。 | 官方来源 · 2026-10-15T07:57:38.509Z |
| AWS NVIDIA L4 x1 aws | GPU · NVIDIA L4 × 1 | — / 24 GB | USD 0.8 / 小时 / 副本 | 按分钟结算 24 GB 为 GPU 显存；官方本表未给出系统内存。实例按分钟结算，容量受配额限制。 | 官方来源 · 2026-10-15T07:57:38.509Z |
| AWS Inferentia2 x1 aws | Accelerator · AWS Inferentia2 × 1 | 14.5 GB / 32 GB | USD 0.75 / 小时 / 副本 | 按分钟结算 32 GB 加速器内存，14.5 GB 系统内存；需兼容推理引擎、可用容量与配额。 | 官方来源 · 2026-10-15T07:57:38.509Z |
小时单价不是 Token 单价。费用由实际计费时长、运行副本与扩缩容决定；资源生命周期规则以官方文档为准。此处不假定整月运行或免费扩容。
专用部署资源按分钟结算，原价按实例小时展示。当前官方要求有效订阅和信用卡；成功部署的端点在初始化和运行时计费，扩容副本的生命周期分别计费。缩容至零需配置最小副本为零，并存在冷启动。区域、容量和配额需在控制台核实；不足一分钟的舍入与最短时长未明确，不生成虚假月费估算。
官方资料冲突与待核验项
- 主定价页 CPU 起价经过舍入，且文档月费示例存在旧计算。采用实例表原始小时单价，不套用旧示例。 /pricing · /docs/inference-endpoints/pricing
- docs/inference-endpoints/index · 2026-10-15T08:05:32.985Z
- docs/inference-endpoints/pricing · 2026-10-15T08:05:32.985Z
- docs/inference-endpoints/guides/autoscaling · 2026-10-15T08:05:32.985Z
Need a recommendation?
The AI advisor considers your budget, tools, usage and purchase conditions to suggest primary and backup options.
Open AI advisor →