GPU RTX 4090
IS-GPU-4090 · 冰岛
- GPU1 × RTX 4090 · 24 GB GDDR6X
- CPU16 vCPU
- 内存64 GB RAM
- 存储1 TB NVMe
起$420/月
订购 IS-GPU-4090Your model, your rules (within the AUP)
Run open-weight models on hardware you control
Hosted AI APIs log prompts, add filters you can't change and retire models on their own schedule. On a GPU server from us, you choose the open-weight model, the system prompt and the settings. Prompts stay on your machine.
适合：AI researchers, developers and privacy-conscious teams who want control of the model and of the data they send it.
本页面目前仅提供英文版本。
适合谁
真正会订购的人
People who can't use a filtered API
Research, fiction, security review and medical drafts get refused by hosted chats for reasons you can't audit. A model you run will answer according to the weights and the prompt you set.
Teams that won't send prompts to a vendor
Customer text, source code and internal notes shouldn't be a training set. Self-hosting means we don't see the prompts, and the API vendor doesn't either.
Builders pinning a version
Hosted model names move. On your GPU the weights change when you change them. That matters for anything you need to reproduce.
包含
你将获得
服务的具体组成。可接受使用政策适用于其中每一项。
No third-party filter in the path
Nothing of ours rewrites the model's output. You install vLLM, llama.cpp, Ollama or your own server and set the policy there.
No prompt logging by us
We don't read, store or train on what you send to your model. We also don't monitor output. We act on credible abuse reports and on valid local legal orders.
Open-weight models
Llama, Mistral, Qwen, DeepSeek and whatever else fits the VRAM. You supply the weights. We supply the card.
Fine-tuning on your data
LoRA and fuller training runs stay on the disk you rent. Backups are your job unless you've arranged one.
Three cards
RTX 4090 24 GB, RTX 5090 32 GB, H100 80 GB. Pick from VRAM, not from a marketing tier name.
The AUP, unchanged
'Uncensored' means we don't filter the model. It does not mean CSAM, malware, phishing or fraud are allowed. You are responsible for how the system is used.
为什么选这里
为什么在 MercureHosting 选择uncensored ai hosting
No third-party filter layer
Nothing sits between you and the model. You decide how it's configured.
No prompt harvesting
We don't look at, log or train on what you send to your own model.
Any open-weight model
Llama, Mistral, Qwen, DeepSeek and others, served with the stack you install.
Fine-tune on your GPU
Train LoRAs and larger runs on your own data, on the card you rented.
Stable versions
The weights stay until you replace them. A vendor can't retire the model out from under a shipped feature.
司法管辖区
GPU 所在地
The model runs in the country of the GPU server: Iceland, Romania, Switzerland or the Netherlands. Pick Switzerland or Iceland if the legal setting matters as much as the card. Pick Romania or the Netherlands if European latency matters more.
North Atlantic · GPU
冰岛
Strong press-freedom tradition, geothermal power, cool air.
~40–60 ms to Western Europe, ~80–100 ms to US East
Eastern Europe (EU) · GPU
罗马尼亚
EU location with some of Europe's fastest consumer networks.
~25–40 ms to Central Europe
Central Europe · GPU
瑞士
Strict data-protection law and a neutral, stable legal system.
~10–25 ms to Western/Central Europe
Western Europe (EU) · GPU
荷兰
Europe's interconnection hub — the lowest latency to most of the continent.
~5–20 ms to most of Western Europe
每个机房页面都列出了优势与取舍。如需候选清单，请使用司法管辖区选择器。
下单
下单流程
所有产品都使用同一个五步表单。账户可选，无需任何身份证件。
- 01
选择产品和国家
VPS、独立服务器、Windows RDP、GPU 服务器或域名。套餐所在国家即该机器适用法律的司法管辖区。
- 02
选择计费周期
月付，或预付 3、6、12 个月。周期越长，实际月价越低。折扣在生成账单前于表单中计算。
- 03
留下一种联系方式
一个邮箱地址或 Telegram 用户名。姓名可选且不核验。表单不要求身份证件、自拍、电话号码或水电账单。
- 04
支付加密货币账单
价格以美元标示。使用 BTC、ETH、USDT、USDC、LTC、DOGE、TRX、SOL、BNB、XRP 或 BCH 为钱包充值。转账到该地址，然后等待工作人员确认。工作人员将转账标记为已付款后余额到账，服务器从该余额中订购。
- 05
获取访问权限
VPS 和 RDP 通常在付款确认后几分钟内就绪。独立服务器和 GPU 服务器可能需要几小时到一天。登录信息会发送到你提供的联系方式。
Order the GPU plan, then install the server stack yourself. A small VPS beside it is enough for an agent or an API gateway if you don't want that process on the GPU box. The agents page describes the planned ordering API. It is not live.
此表单的周期折扣：季付 −25% · 半年付 −35% · 年付 −50%。
对比
与典型 KYC 主机商并列
这里的典型主机商是指在出租服务器前要求政府证件或已验证银行卡的服务商。各公司情况不同，决定前请阅读其最新条款。
| 项目 | MercureHosting | 典型 KYC 主机商 |
|---|---|---|
| Who filters output | You do, in the software you install. We don't add a filter. | The API vendor, with a policy you can't read line by line. |
| Prompt data | Stays on your server. We don't train on it. | Logged under their privacy policy, sometimes used to improve models. |
| Model lifetime | The file on your disk, until you delete it. | A model name they can deprecate. |
| Account check | Email or Telegram and a crypto invoice. | A platform account, often with organisation verification for higher limits. |
| Illegal use | Still prohibited by the AUP and by local law. | Also prohibited, plus their content policy. |
套餐
合适的套餐
以美元标价，以加密货币结算。订单表单会在付款前应用周期折扣。
GPU RTX 5090
IS-GPU-5090 · 冰岛
- GPU1 × RTX 5090 · 32 GB GDDR7
- CPU24 vCPU
- 内存96 GB RAM
- 存储2 TB NVMe
起$690/月
订购 IS-GPU-5090GPU H100
IS-GPU-H100 · 冰岛
- GPU1 × H100 80 GB (PCIe)
- CPU32 vCPU
- 内存192 GB RAM
- 存储4 TB NVMe
起$2,700/月
订购 IS-GPU-H100价格以美元计，以加密货币结算。周期越长最高可省 50%。
常见问题
关于这点我们常被问到
What does 'uncensored' mean here?
No third party filters or rewrites what your model says. You are responsible for how you use it, and the AUP still applies, including the ban on CSAM and on fraud.
Do you monitor model output?
No. We don't inspect workloads. We act on credible abuse reports and on valid legal orders from the hosting country.
Which GPU do I need?
A 24–32 GB card handles most quantised 7B–70B models. Large unquantised models and heavier training need H100-class hardware. The GPU guide goes into more detail without inventing benchmark scores.
Can I serve an API to my own users?
Yes. You run the HTTP server. Rate-limit it, and don't let it become an open relay for abuse. The token owner is responsible for what the endpoint does.
Do you provide the model weights?
No. You download the open-weight model you have the right to use. We provide the GPU server, drivers and JupyterLab.
Will you ban a jailbreak?
A jailbreak of a model you run is your configuration. Use of the server for activity in the AUP, such as malware or fraud, is what gets a service suspended.
可接受使用政策适用于所有服务
隐私不是滥用的许可。CSAM、恶意软件、网络钓鱼、垃圾信息、僵尸网络、欺诈以及在服务器所在地违法的任何内容均被禁止，违反这些规则的服务将被暂停。我们遵守服务器所在司法管辖区发出的有效法院命令。请阅读完整的 服务条款与可接受使用政策.
Put the model on your own GPU
下单只需一个邮箱或 Telegram 用户名和一笔加密货币付款。
每个页面都有
离岸 VPS 一目了然
离岸 VPS 是一台位于你所选国家的 KVM 服务器，用邮箱或 Telegram 用户名下单，以加密货币付款。国家决定了机器适用的法律。来自其他地方的投诉不会导致自动下架，可接受使用政策在七个机房均适用。
免 KYC
只需一个邮箱地址或 Telegram 用户名即可下单。我们不要求护照、自拍、电话号码或地址证明。姓名可选，且不做核验。
加密货币结算
使用 BTC, ETH, USDT, USDC, LTC, DOGE, TRX, SOL, BNB, XRP, BCH 付款。价格以美元标示。先在钱包中充值，再用余额下单。无银行卡，无银行账户。转账到该地址，然后等待工作人员确认。
七个司法管辖区
冰岛、巴拿马、摩尔多瓦、罗马尼亚、瑞士、荷兰、俄罗斯。你在下单时选择国家，任何投诉都依据该国法律审理。在你支付长期费用之前，各机房页面会讲清利弊。
有明确清单的隐私
我们不记录 VPS 收发的流量，也不保存磁盘内容。存在的计费记录包括：一种联系方式、订单、付款凭证、IP 和电源状态。隐私政策列出了完整清单。
坦诚处理通知
美国 DMCA 通知不会导致自动下架。投诉依据 VPS 所在国家的法律审核。该国的有效法院命令会被执行。依据该国法律构成的版权侵权仍然违反可接受使用政策。
KVM、root 与 VNC
每台 VPS 都是 KVM，拥有完整 root，网络中断时还有 VNC 控制台。可安装模板或上传自己的 ISO。这不是被面板挡在中间的容器。
VPS 本身还包括：约 3,500 MB/s 级别的 NVMe、合理使用的不限流量端口、季付 −25%、半年付 −35%、年付 −50%，以及与典型 KYC 主机商的对比。我们不公布实测可用率，周期折扣之外也没有优惠码。