Large open models without your own hardware: what privacy you really get
Some people want to try a strong uncensored model. That does not make them criminals. Most of them simply do not want to be watched: they do not want every question logged, stored, and maybe used for training. Others are tired of a model that refuses a harmless question about security, medicine or fiction.
The clean answer is a local model on your own box. But the biggest open models do not fit on a home machine. My GB10 box has about 121 GiB of usable memory, and the models below need 300 to 600 GB. So this note is about the non-local options. There are two:
- A hosted service that already runs uncensored models.
- A rented GPU node, where you run the weights yourself.
One thing first, so nobody reads this the wrong way: privacy in both options is conditional. You send your prompts to a machine that belongs to someone else. A "no logs" service can log. Any of these services could even be a honeypot, built to collect exactly the questions people do not want to ask elsewhere. I have no evidence that any of them is, but from the outside you cannot rule it out. Only a local model on your own hardware can guarantee privacy, and even that only if the machine is offline. Inference servers and their Python or Node.js dependencies can send telemetry or update themselves, and "safe" packages in public repositories have been hijacked before. For real privacy, run the model on an air-gapped box, or at least block all of its outgoing traffic. Everything below is a trade: less privacy for more model.
I have not benchmarked the services or the models below. Prices and privacy statements are taken from the providers' own pages on 20 November 2026, unless I say otherwise. Check them before you pay.
Contents
What "uncensored" means here
Almost all of these models are abliterated: someone found the direction in the weights that produces refusals and removed it. The model is not retrained and it does not become smarter. It just stops saying no. The law does not change either: what was illegal to do stays illegal.
Option 1: hosted services
The fastest way. You pay per token or per month and you need no hardware.
| Service | What you get | Price | What they say about your prompts |
|---|---|---|---|
| NanoGPT | Chat and API, about 600 text models, several uncensored ones (GLM 5.3 Uncensored, DeepSeek V4.1 Flash Uncensored, Abliterated Large) | Pay as you go, deposit from $0.10 in crypto or $1 by card. PRO is $12 per month. GLM 5.3 Flash Uncensored is $0.20 / $0.80 per 1M tokens | "By default, we do NOT store prompt or conversation content on our servers." No training on your content. Works without an account |
| Venice | Chat and API, own model "venice-uncensored" | Free tier (10 text prompts a day, no card), Pro $18 per month, or pay per token | "Zero data retention by default" on the API. Says illegal content is blocked at every setting |
| Featherless | API and chat, hundreds of abliterated open models, mostly small and mid-size | Chat plan $25 per month, developer plan from $50 per month | "Featherless does not log chats, prompts, or completions sent through our API." |
| abliteration.ai | API, one large abliterated model | Plans from $20 per month or pay as you go; about $3.00 / $5.00 per 1M tokens according to a third-party comparison | "Your prompts and responses stay out of our logs." |
| OpenRouter | A router to other providers, includes Venice Uncensored | Per token, plus a card fee | Logging only if you opt in, but each upstream provider has its own policy |
Things to keep in mind:
- "No logs" is a promise, not a proof. You cannot verify it from outside, and a service that advertises "uncensored and private" is also the perfect honeypot. Do not type anything there that you could not afford to see leaked.
- Routers pass your prompt on. NanoGPT and OpenRouter send the request to a third-party host. NanoGPT's own policy calls "zero data retention" a best-effort label based on what that host claims. Read the privacy page and its settings before you rely on it.
- Payment can identify you. A card does. Crypto is visible on the chain.
- Start small. NanoGPT and Venice let you try with no account or no card. Subscribe only if the answers are good enough.
Option 2: rent a GPU node and run the weights yourself
More work, but your prompts go only to a machine you rent for a few hours. That is better than a chat service, not the same as private.
Where to get the models
Hugging Face, filtered to GGUF text models with more than 500B parameters and "uncensored" in the name: the search. Five from the top of that list:
| Model | Size | Method | File | Notes from the model card |
|---|---|---|---|---|
| AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-GGUF | 552B | abliteration | IQ1_M to IQ2_XS, 121 to 155 GB in the file list | Needs the Strata runtime, not plain llama.cpp. MIT |
| AtomicChat/DeepSeek-V4.1-Flash-1017-Uncensored-DwarfStar-GGUF | 755B | abliteration | Q2_K, 366 GB | Needs the DwarfStar runtime. MIT |
| softwareweaver/GLM-5.3-Uncensored-GGUF | 753B | converted from an uncensored FP8 release | Q4_K_M, 455 GB | Runs with llama.cpp. MIT |
| Ryanchen911/Kimi-K3-Uncensored-GGUF | 2.8T | abliteration | IQ1_S, 580 GB | Needs a llama.cpp fork and about 550 GB of RAM plus VRAM. Kimi K3 license |
| Harsha001/DeepSeek-V4.1-Flash-1017-Uncensored-GSQ-RCO-StrixHalo-GGUF | 658B | pruned and abliterated | 215 GB | Made for one AMD machine with 128 GB of unified memory, not for a rented NVIDIA node. MIT |
Two warnings before you download half a terabyte:
- Read the card. Three of these five need a special runtime or a fork. The GLM 5.3 file is the only one that the card says runs with normal llama.cpp, so it is the easy first try.
- Low-bit files lose quality. Most of these are 1-bit and 2-bit quants. In my own test a 30B model broke below 4 bits. Very large models are said to survive it better, but I have not measured these. The 4-bit GLM file is the safer bet.
How much memory you need
Rough sizing: memory = parameters x bytes per parameter, plus 10 to 20% for the KV cache. BF16 is 2 bytes per parameter, FP8 is 1, 4-bit is about 0.5 to 0.6. For a GGUF file it is simpler: the file size plus the cache.
| H200s in one node | Memory | Fits roughly |
|---|---|---|
| 1 | 141 GB | 70B in FP8, about 200B in 4-bit |
| 2 | 282 GB | 70B in BF16, about 400B in 4-bit |
| 4 | 564 GB | about 400B in FP8; the 455 GB GLM 5.3 file |
| 8 | 1128 GB | 600 to 700B in FP8, the largest open models |
Rent the GPUs in one node (2, 4 or 8 on the same machine). Separate single-GPU instances cannot be joined. Many providers sell H200 only as 8-GPU nodes.
How the model is split depends on the format:
- GGUF with llama.cpp: it spreads the layers over all GPUs it sees, and what does not fit can stay in system RAM (slower).
- Safetensors (BF16, FP8) with vLLM or SGLang: tensor parallelism, one
flag, for example
--tensor-parallel-size 8.
What it costs
All prices are per GPU per hour, also where the provider rents only whole nodes. Overseas prices come from aggregators, Chinese prices from rentgpu.link, both November 2026. CNY is converted at 7.25 per dollar. They change daily.
H200, 141 GB
| Provider | Price per GPU-hour | Type |
|---|---|---|
| Vast.ai | from $0.74 | spot, marketplace |
| Nebius | $0.79 | spot |
| Beam | $2.09 | on demand |
| Together AI | $2.99 | on demand |
| Koyeb | $3.00 | on demand |
| DigitalOcean | $3.44 to $4.47 | on demand, sources differ |
| RunPod | $3.59 to $5.29 | on demand, sources differ |
| Gongji Compute (China) | 35.00 CNY, about $4.83 | pay as you go |
| AWS | about $10.29 | on demand, only as an 8-GPU node (p5e.48xlarge, $82.29 per hour) |
A 4x H200 node is about $8 to $12 per hour on demand, an 8x node $16 to $32 at the cheaper providers, and less on spot. AWS is the expensive end: its 8x H200 node is $82.29 per hour on demand in us-east-1. Reserved in advance as a Capacity Block, the same node cost $39.80 per hour after a price rise in January 2026.
Count the download too: 455 GB takes about an hour at 1 Gbit/s, and on a GPU node you pay for that hour. Download the model once to cheaper storage instead: on AWS into S3 from a small CPU instance (incoming traffic is free, and so is the transfer from S3 to an instance in the same region), on other providers into a storage volume that you attach to the GPU node later. You then pay for storage, not for GPU hours spent waiting.
China, other cards (cheapest listing per card)
| GPU | Memory | Price per card-hour | Platform |
|---|---|---|---|
| H20 | 96 GB | 3.40 CNY, about $0.47 | PPIO, spot |
| Ascend 910B2 | 64 GB | 2.97 CNY, about $0.41 | JiZhiSuan |
| A800 | 80 GB | 5.15 CNY, about $0.71 | MornAI |
| A100 | 80 GB | 5.98 CNY, about $0.82 | AutoDL |
| H800 | 80 GB | 10.50 CNY, about $1.45 | AutoDL |
| H100 | 80 GB | 11.88 CNY, about $1.64 | PPIO, spot |
| Ascend 910C | 128 GB | 13.20 CNY, about $1.82 | Gongji Compute |
Notes on China:
- The H20 is the cheapest large memory on the list. It is slow at compute, but for inference of a big model memory matters most.
- Ascend cards are not CUDA. They use Huawei's own stack, CANN. llama.cpp has a CANN backend, and projects like cuda-morph try to run CUDA-style PyTorch code on Ascend, but cuda-morph is alpha and its Ascend support "needs hardware validation". Expect extra work.
- Registration may need Alipay or a Chinese bank card. Alibaba Cloud International takes normal cards, but its big-cloud prices are higher.
- Your prompts and files sit under Chinese jurisdiction.
Other hardware you can rent
- A large CPU node. Plenty of RAM instead of GPU memory. For example, AWS
r5.24xlargehas 96 vCPUs and 768 GiB of RAM, enough for the 455 GB GLM 5.3 file. Spot prices for such instances can be very low in some regions; check the spot pricing page. Expect it to be much slower than GPUs, because every token has to read the weights from ordinary RAM. Mixture-of-experts models suffer less, since they read only a part of the weights per token. I have not measured it. - Dedicated LLM hosting. For example HOSTKEY: GPU servers with Ollama and Open WebUI pre-installed, models up to gpt-oss:120b, from 0.624 EUR per hour (RTX 5090) or 299 EUR per month (RTX 4090). Locations named on the page: the Netherlands, the USA and Russia. The page says nothing about logging. Easy to start, but it is the provider's setup, not yours, so assume some tracking.
- AWS Trainium2. A
trn2.48xlargehas 16 Trainium2 chips with 1.5 TB of accelerator memory, enough for the largest open models. AWS claims 30 to 40% better price-performance than its H200 instances; on demand it is $103.16 per hour. It is not CUDA: it needs software built on the AWS Neuron SDK, so the GGUF files above are not the way to use it. - The big clouds (AWS, Google Cloud, Azure): mature and reliable, but GPU instances on demand are the most expensive; spot is cheaper. See AWS on-demand pricing.
- Ordinary web hosting, VPS or dedicated servers: technically possible, but many hosters treat sustained heavy compute as abuse, even on your own dedicated machine. Read the terms first.
Is a rented node private?
More than a chat service, less than your own box. Nobody logs your prompts by design, but the provider owns the machine and can technically look inside. On a marketplace like Vast.ai the host is some third party you know nothing about. Use a known provider if the content matters, and delete the instance and its disk when you are done.
Your own hardware does not have to be a DGX or a Mac
For the "real privacy" option, the hyped choices are DGX boxes, Mac Studio and other machines with large unified LPDDR memory. There are cheaper ways to get a lot of memory:
- A desktop PC with 256 GB of ordinary DDR4 or DDR5. Slow, because the weights are read from RAM, but enough for a mid-size model in 4-bit or a mixture-of-experts model.
- Unlocked NVIDIA CMP 170HX mining cards. Since August 2026 a tool restores the hidden memory: 8 GB cards become 64 GB, 10 GB cards 80 GB, on chips with good memory. Prices on eBay jumped from about $100 to $200 to $1,200 to $2,000. Stability at 64 GB is not guaranteed, the bus is PCIe Gen2 x4 and there is no FP8.
- Modded RTX 4090 with 48 GB. About $3,400. Loud, up to 65 dB in one test, and no warranty.
I have not tried any of these.
Which one to pick
- Just curious: a hosted service. Try the free tier or put in one dollar.
- You want the largest model, or you do not trust any "no logs" promise: rent a 4x or 8x H200 node for an evening. That is $30 to $100.
- You need real privacy: a smaller model on your own hardware. It is the only setup where nobody else is in the loop, and the only one that can guarantee anything.
Changes
- 21 November 2026. All prices in the H200 table are per GPU-hour; the AWS row was corrected to the p5e.48xlarge on-demand price. Added: download the model once to S3 or a storage volume, llama.cpp and cuda-morph for Ascend, AWS Trainium2, and cheaper own hardware (256 GB desktop, unlocked CMP 170HX, 48 GB RTX 4090). Suggestions by Shelwien.
- 20 November 2026. First version.
Sources
- NanoGPT models, pricing, privacy policy
- Venice: best uncensored LLM APIs
- Featherless: best uncensored AI models 2026
- Brutal AI: uncensored LLM API providers compared
- Off Grid AI: best private LLM APIs in 2026
- flopper.io: H200 cloud pricing
- Thunder Compute: H200 price comparison
- vps.fund: H200 cloud price comparison
- rentgpu.link: GPU prices on Chinese platforms
- PromptQuorum: AutoDL, Alibaba and Tencent GPU pricing
- HOSTKEY: LLM hosting
- AWS EC2 on-demand pricing, AWS EC2 spot pricing
- Vantage: p5e.48xlarge, Vantage: trn2.48xlarge
- The Register: AWS raises GPU prices 15%
- AWS Trn2 instances
- AWS S3 pricing
- cuda-morph on PyPI
- Wccftech: CMP 170HX unlocked to 64/80 GB
- Igor's Lab: modded RTX 4090 with 48 GB
Thanks to Shelwien for the review.