An FHE coprocessor network for language models.
Developers send encrypted prompts and receive encrypted outputs. Independent node operators run the encrypted computation on accelerated hardware and are paid per job. The encryption keys never leave the user.
An API that feels normal
SDK and API that look like a regular AI inference API, with encryption and decryption handled on the client.
GPUs become coprocessors
Software that turns GPU servers, and later dedicated FHE accelerators, into paid coprocessors.
"We can't see your data"
AI features where privacy is a technical fact, not a promise.
Key generation
The SDK creates a secret key on your device and a public evaluation key that lets servers compute on ciphertext.
READABLE · YOUEncryption
The prompt is tokenized and embedded locally, then encrypted.
CIPHERTEXTEncrypted inference
A CIPHRA node runs the model's layers on ciphertext. Encrypted numbers in, encrypted numbers out.
CIPHERTEXTVerification
The client checks the agreed model was run. At launch: redundant execution on independent nodes with matching results.
CIPHERTEXTDecryption
Only your secret key turns the result into readable text.
READABLE · YOUEvery prompt is a confession.
A doctor pastes a scan report into an assistant. A lawyer checks a draft settlement. In almost every AI product today, that confession arrives at someone else's server in plain text. The organizations with the most valuable data are not afraid of the model. They are afraid of the moment their data becomes readable.
What if the model could answer without ever being able to read the question?
| Option | What it protects | What it leaves open |
|---|---|---|
| Standard AI API + privacy policy | Legal promises, contracts | Data is readable by the provider and its infrastructure |
| Self-hosting on-premise | Data stays inside the company | Expensive hardware, limited model quality, heavy operations |
| Trusted execution (TEE) | Data isolated inside secure hardware | Trust in the chip vendor; history of side-channel attacks |
| Anonymizing before sending | Removes obvious identifiers | Context can re-identify people; useful detail is lost |
| > CIPHRA · FHE | Data protected by mathematics, not by trusting the operator, vendor or a policy | Speed, which is exactly what CIPHRA focuses on |
Peer-reviewed math. Accelerated hardware.
CIPHRA builds on established FHE schemes instead of inventing new cryptography, on top of mature open-source libraries such as OpenFHE and TFHE-rs. Every CIPHRA-specific component will be open-sourced and audited.
Approximate arithmetic on real numbers
Suits the matrix multiplications that dominate neural networks.
Fast non-linear functions
Programmable bootstrapping evaluates non-linear steps quickly. The CIPHRA compiler chooses the scheme per layer.
OpenFHE · TFHE-rs
Mature open-source foundations. CIPHRA-specific components will be open-sourced and audited.
Polynomial approximation
Softmax, GELU and normalization approximated per model, with accuracy loss kept within a stated tolerance.
Hybrid splitting
Your device runs tokenization, embedding and decoding. The network runs the heavy middle layers encrypted.
FHE-aware fine-tuning
Open-weight models are tuned so they keep their quality after approximation.
Data-center GPUs
Custom CUDA kernels for NTT and key switching, on hardware node operators already own.
FPGA cards
Lower latency and better energy use for production workloads.
FHE ASICs
Dedicated accelerator chips, sourced through hardware partnerships as they reach the market.
// performance, stated honestly: encrypted inference is still much slower and more expensive than plain inference. First targets are small and mid-size models: classification, embeddings for private search, document extraction and compact language models. CIPHRA will publish latency, cost per request and accuracy against the unencrypted baseline for every supported model.
Five things you can build on.
Patient records, unexposed
Triage, coding and summarization without exposing records to the AI provider.
Privileged documents
Contract review and e-discovery on documents that cannot leak.
Data that can't leave the bank
Fraud scoring and risk analysis on customer data that stays unreadable.
Assistants that can't peek
Internal assistants read sensitive documents without the vendor seeing them.
Private strategies
AI agents act on private inputs, like a trading strategy, without revealing them to the network.
Build it on the testnet
Developer testnet ships in Phase 1. Join the whitelist to hear first.
$CPH is built to be used.
settles in $CPH.
- name
- CIPHRA
- type
- Utility token of the CIPHRA network
- launch
- Through a launchpad. Supply, allocation, vesting and sale terms are set and published by the launchpad.
- contract
- TBA · announced at launch
- x
- @CIPHRA_FHE
- telegram
- t.me/CIPHRA_FHE
Pay for compute
Encrypted inference jobs are priced in $CPH. Enterprises that want stable pricing can pay in stablecoins, converted at settlement.
Node staking
Operators stake $CPH to receive jobs. A node whose results fail verification loses part of its stake.
Priority & capacity
Stakers can reserve guaranteed capacity for production workloads.
Governance
Holders vote on supported models, fee parameters, grants and protocol upgrades.
From proof to scale.
Proof
- Open-source compiler alpha
- Encrypted classification and embedding models on GPU
- Public benchmarks
- Developer testnet
Private RAG
- Private RAG and Inference API beta
- First external node operators
- Security audit of SDK and node software
- $CPH launch on launchpad
Language models
- Encrypted inference for compact LLMs with hybrid splitting
- FPGA nodes
- Enterprise pilots in healthcare and finance
Scale
- Larger models as accelerators mature
- ASIC partnerships
- Full on-chain governance
// milestones are targets and depend on funding, audits and hardware availability.
Security model, honest limits.
Computed on.
Never read.
Against node operators, CIPHRA and network observers
- Prompt content stays encrypted end to end
- Model outputs are only readable with your key
- Retrieved documents in Private RAG stay encrypted
What FHE does not cover
- Metadata: that a request was made, its size, timing and which model was called
- Model weights: known to nodes in the initial design
- Key safety: a compromised device or key can't be protected by encryption
- Accuracy: approximated models may differ slightly; each ships with a published accuracy report
Seal your spot on the whitelist.
Drop your EVM wallet, add your X handle, then let CIPHRA write your launch post. Your address is shown encrypted below as you type, just for fun. It is stored only for the whitelist.
Spots are sealed in order. One wallet, one spot.
Read the whitepaper.
- 01Every prompt is a confession
- 02Why today's options fall short
- 03What CIPHRA is
- 04How it works
- 05The technology
- 06Products
- 07Who it is for
- 08The $CPH token
- 09Roadmap
- 10Security model and honest limits
- 11Risks and disclaimer
