Zero data retention LLM API: what is kept, what is not, and which models
On Heabsy's own models — 12 of them, including Qwen3.8 27B, GLM 5.3 and MiMo V2.6 Flash — prompts and completions are processed in memory on GPUs in the European Economic Area and are never written to disk or logs. What we keep is billing metadata: model, token counts and time, no text. The 23 partner-hosted models in the same catalog, such as Claude and GPT-6, do not have zero data retention, and this page says which is which.
Create a keyModels with zero data retention
Generated from the catalog at build time. Every model below runs in the EEA and stores no prompt or completion text.
| Model | ID | Context | Input / 1M | Output / 1M |
|---|---|---|---|---|
| Qwen3.8 27B Fable | qwen38-fable | 512K | $0.095 | $0.58 |
| Qwen3.8 27B Queen | qwen38-queen | 512K | $0.095 | $0.58 |
| GLM 5.3 Flash Uncensored | glm53-flash-unc | 1M | $0.1739 | $0.6957 |
| MiMo V2.6 Flash Uncensored | mimo-v26-flash-unc | 1M | $0.4404 | $1.4285 |
| MiMo V2.6 Flash | mimo-v26-flash | 1M | $0.4762 | $1.4286 |
| MiMo V2.6 Flash Abliterated | mimo-v26-flash-abl | 1M | $0.4762 | $1.4286 |
| GLM 5.3 | glm53 | 1M | $0.5217 | $0.9565 |
| Qwen3.8 27B Uncensored Fast by Heabsy | qwen38-unc-fast | 512K | $0.7904 | $2.4952 |
| Qwen3.8 27B Uncensored | qwen38-unc | 1M | $0.8095 | $2.5238 |
| Qwen3.8 27B | qwen38 | 256K | $1.00 | $3.00 |
| GLM 5.3 Uncensored | glm-5.3-uncensored | 1M | $1.1905 | $2.1429 |
| CyberHeabsy | cyberheabsy | 1M | $1.4286 | $4.2857 |
Models without zero data retention
Partner-hosted, processed by third-party inference providers, not guaranteed in the EEA:Claude Opus 5, Claude Opus 5.5, Claude Sonnet 5, Claude Sonnet 5.5, DeepSeek V3.2, DeepSeek V4 Flash, DeepSeek V4 Pro, DeepSeek V4.1 Flash, GLM 5.2, GLM 5.3, GLM 5.3 Flash, GPT-5.6 Luna, GPT-5.6 Sol, GPT-6 Astra, GPT-6 Luna, GPT-6 Sol, GPT-6.1 Sol, gpt-oss-120b, Kimi K3, MiMo V2.6 Pro, MiniMax M3, Qwen3.8 Flash, Tencent Hy4 Preview. Use them for data where that is acceptable.
What is kept, and for how long
| Data | Retention |
|---|---|
| Prompt and completion text, own models | Not persisted. Processed in memory for the duration of the request; a prefix cache in GPU memory speeds up the next turn of the same conversation and is not written anywhere. |
| Usage metadata and financial records | 10 years, as accounting law requires: model, token counts, time, status, the key that made the request. No text. |
| Application logs | 30 days, without prompt or completion content. |
| Security audit logs | 1 year. |
| Prompt and completion text, partner-hosted models | Not stored by Heabsy; the third-party provider retains it under its own policy — no zero data retention. |
The binding version is the retention table in our privacy policy.
Four promises that are often mixed up
"GDPR-compliant LLM API" usually stands for one or two of these. They are independent: a provider can be an EU company and still keep your prompts, or keep nothing and still process outside Europe.
| Promise | What it means | At Heabsy |
|---|---|---|
| Zero data retention | Prompt and completion text is not written to disk or logs; it exists in memory while the request runs. | Our own models (12). Not partner-hosted models. |
| Processing in the EEA | The GPU that runs the request is inside the European Economic Area. | Our own models. Partner-hosted models: not guaranteed. |
| EU contracting party | Your contract, invoice and data processing agreement are with a company in the EU. | All models — one key, one invoice. |
| No training on your data | Inputs and outputs are not used to train or fine-tune models. | We do not train on customer content; for partner-hosted models the provider's policy applies. |
How to check it yourself
- The public catalog
GET https://api.heabsy.com/site/v1/modelslists every model withdeployment_regionandcompliance.zdr— the same data this page is built from. - The privacy policy lists the sub-processors and where they process; the data processing agreement sets the terms.
- Restrict a key to the models above in the console: requests for any other model are refused, so nobody on your team sends customer data to a partner-hosted model by mistake.
When Heabsy is not the right choice
- You need Claude, GPT or Gemini themselves with processing in Europe or zero retention. Our partner-hosted versions have neither; the EU regions of the large cloud providers are the route for that, and so is OpenRouter's EU routing — compared on OpenRouter alternative.
- Your procurement requires a specific audit report or certificate from the API provider.
- You need confidential computing with hardware attestation.
- You need a contractual uptime SLA or a dedicated cluster — that is a separate agreement, not the self-service API.
FAQ
Does Claude Opus 5.5 have zero data retention at Heabsy?+
No. Claude models in our catalog are partner-hosted: the request goes to a third-party inference provider, without zero data retention and without an EEA guarantee. The same applies to GPT-6, Kimi K3 and DeepSeek V4.
Does GLM 5.3 have zero data retention?+
Our own GLM 5.3, ID glm53, does: it runs in the EEA and its prompts are not stored. The catalog also has a partner-hosted GLM 5.3, ID glm-5.3, which does not. The titles of their pages say which is which: (EEA) or (partner-hosted).
Does prompt caching break zero data retention?+
No. The cache is a prefix of the conversation kept in GPU memory on the node that serves it, so the next turn does not recompute it. It is never written to disk and never served to another customer.
Is zero data retention enough for GDPR?+
It removes one risk — stored conversation text — but not your obligations as controller: a lawful basis, a data processing agreement, a DPIA where required, and minimising what you send. Our data processing agreement is public; we do not promise that your use is compliant.
How can I check which models have ZDR without trusting this page?+
The public catalog at api.heabsy.com/site/v1/models carries deployment_region and compliance.zdr for every model, and the retention table in our privacy policy lists what is kept and for how long.