BenchmarksStack Ranking
APIsPricingDocsWhite PaperTokenBlogAboutSecurity Demo
Log InGet API Key
PRODUCTION · 2.21M AUTH/SEC

1.595 Million Authentications Per Second.
Zero Data Exposure.

BFV lattice-based fully homomorphic encryption. N=4096 polynomial ring, 56-bit modulus, SIMD batching. 32 users verified in a single 1,109µs ciphertext operation.

0µs
Per Auth
0/s
Sustained
0
Users/CT
0-bit
NIST L1

Single API Call, Fully Post-Quantum

Three stages execute in sequence. One API call. Every component is lattice-based or hash-based. No classical-only dependencies.

1 FHE Batch
BFV Inner Product
Encrypted biometric matching over 32 users simultaneously. SIMD-packed ciphertext, NTT-domain fused inner product with single final INTT.
32 users/CT N=4096 BFV Lattice
~1,109µs
2 ZKP
STARK Lookup
Zero-knowledge proof that the FHE computation was performed correctly. Cached STARK with SHA3-256 hash commitment. Sub-microsecond verification.
SHA3-256 STARK Cached Hash-based
~0.067µs
3 Attestation
SHA3 + Dilithium
SHA3 digest of the entire computation, then a single Dilithium (ML-DSA) sign+verify to attest the batch. Post-quantum attestation per 32-user batch.
ML-DSA FIPS 204 Batch Lattice
~240µs
Total Pipeline Latency
~1,615µs / 32 users
~42µs per authentication

Engineering That Ships at Scale

Every component is optimized for production throughput on Graviton4 ARM. No research prototypes. No toy demos. Real numbers, real hardware.

967µs Full Pipeline
Encrypt, prove, attest in under 1ms. 32 users per batch with SIMD packing. Fastest production FHE on the market.
STARK Zero-Knowledge Proofs
SHA3-256 based, no trusted setup. First proof: 2.0µs. Cached lookups via DashMap: 0.062µs. Post-quantum by construction.
ML-DSA-65 Attestation
Every batch is signed with Dilithium Level 3. Verifiable proof of correct computation. Non-repudiable audit trail.
SIMD Batching (32 Users/CT)
4096 polynomial slots ÷ 128 biometric dimensions = 32 users per ciphertext. Constant time regardless of batch fill.
Montgomery NTT
Hardware-optimized Number Theoretic Transform with Harvey lazy reduction. Zero division in the hot path. Graviton4 NEON accelerated.
256KB Templates
128× reduction from raw biometric storage. CRT encoding with t=65537. Encrypted templates stored in NTT domain for zero-overhead matching.

Built for Production Privacy

H33-128 BFV enables computation on encrypted data at scale. The data never leaves encryption — not during transit, not during processing, not ever.

Biometric Authentication
Encrypted 128-dim face/fingerprint vectors. Match against enrolled templates without decryption. 32 users verified per batch in ~967µs.
Financial Transaction Screening
Encrypted KYC/AML checks against watchlists. FHE inner product scoring on encrypted identifiers. Sub-millisecond per batch.
Healthcare Identity Verification
HIPAA-compliant patient matching. Biometric templates never leave ciphertext. Dilithium audit trail for compliance.
Access Control & MFA
Encrypted biometric MFA for high-security environments. Zero-knowledge proof of identity. No biometric database to breach.

SECURITY

NIST Level 1 BFV Parameters

Conservative parameter selection validated against the Homomorphic Encryption Standard. All operations in the polynomial ring Z[X]/(XN+1).

Parameter Value
Polynomial degree (N) 4,096
Ciphertext modulus (Q) 56-bit single
Plaintext modulus (t) 65,537
Standard deviation (σ) 3.2
SIMD slots 4,096 ÷ 128 = 32 per ciphertext
Security level 128-bit (NIST Level 1)

Three Tiers of Cryptographic Trust

Choose the trust level your application requires. Each tier builds on the one below it. All tiers include full FHE computation.

Tier 1: FHE Compute
Raw BFV homomorphic encryption operations. Key generation, encrypt/decrypt, arithmetic, and biometric matching. No attestation layer — you manage trust externally.
POST/v1/fhe/h33-128/keygen
POST/v1/fhe/h33-128/encrypt
POST/v1/fhe/h33-128/decrypt
POST/v1/fhe/h33-128/add
POST/v1/fhe/h33-128/multiply
POST/v1/fhe/h33-128/inner-product
POST/v1/fhe/h33-128/biometric-match
GET/v1/fhe/h33-128/health
Example
curl -X POST https://api.h33.ai/v1/fhe/h33-128/encrypt \
  -H "Authorization: Bearer $H33_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"plaintext": [1, 0, 1, 0, ...], "public_key_id": "pk_abc123"}'
Tier 2: FHE + Dilithium Attestation
Everything in Tier 1, plus post-quantum Dilithium (ML-DSA) signatures on every compute result. All FHE endpoints automatically return a dilithium_signature field.
POST/v1/fhe/h33-128/keygen
POST/v1/fhe/h33-128/encrypt
POST/v1/fhe/h33-128/decrypt
POST/v1/fhe/h33-128/add
POST/v1/fhe/h33-128/multiply
POST/v1/fhe/h33-128/inner-product
POST/v1/fhe/h33-128/biometric-match
GET/v1/fhe/h33-128/health
POST/v1/fhe/h33-128/dilithium/keygen
POST/v1/fhe/h33-128/dilithium/sign
POST/v1/fhe/h33-128/dilithium/verify
Example — Every compute response includes attestation
curl -X POST https://api.h33.ai/v1/fhe/h33-128/biometric-match \
  -H "Authorization: Bearer $H33_API_KEY" \
  -H "X-H33-Tier: dilithium" \
  -d '{"ciphertext_id": "ct_xyz", "template_id": "tmpl_abc"}'

# Response includes:
# "dilithium_signature": "base64...",
# "dilithium_public_key": "base64..."
Tier 3: FHE + H33-3-Key Triple Attestation
Maximum trust. Every compute result carries a full H33-3-Key nested signature bundle: Ed25519 + Dilithium-5 + FALCON-512. Three independent mathematical families. Zero single points of failure.
POST/v1/fhe/h33-128/keygen
POST/v1/fhe/h33-128/encrypt
POST/v1/fhe/h33-128/decrypt
POST/v1/fhe/h33-128/add
POST/v1/fhe/h33-128/multiply
POST/v1/fhe/h33-128/inner-product
POST/v1/fhe/h33-128/biometric-match
GET/v1/fhe/h33-128/health
POST/v1/fhe/h33-128/dilithium/keygen
POST/v1/fhe/h33-128/dilithium/sign
POST/v1/fhe/h33-128/dilithium/verify
POST/v1/fhe/h33-128/3key/keygen
POST/v1/fhe/h33-128/3key/sign
POST/v1/fhe/h33-128/3key/verify
Example — Full 3-layer signature bundle on every response
curl -X POST https://api.h33.ai/v1/fhe/h33-128/biometric-match \
  -H "Authorization: Bearer $H33_API_KEY" \
  -H "X-H33-Tier: 3key" \
  -d '{"ciphertext_id": "ct_xyz", "template_id": "tmpl_abc"}'

# Response includes full 3-layer signature bundle:
# "h33_3key_signature": "base64...",
# "ed25519_pk": "base64...",
# "dilithium_pk": "base64...",
# "falcon_pk": "base64..."

Included with every API call
H33 Noise Pilot

Automatic Noise Management. Zero Configuration.

Every H33-128 API call is powered by Noise Pilot — our AI-driven noise budget optimizer. It tracks, schedules, and optimizes FHE noise automatically so you never worry about ciphertext corruption or budget exhaustion.

Noise Budget Tracking
Every operation returns remaining noise budget in bits. Noise Pilot monitors degradation in real-time and warns before exhaustion.
Auto Mod-Switch Scheduling
Noise Pilot automatically schedules modulus switching before budget exhaustion, extending computational depth without developer intervention.
AI Crypto Health — 0.52µs
Native Rust AI agent continuously assesses parameter health, noise distribution, and operational integrity at sub-microsecond latency.
Operation Recommendations
Query the Noise Pilot API for optimal operation ordering, parameter selection, and depth estimates before executing your FHE pipeline.

Frequently Asked Questions

What is homomorphic encryption in plain English?
Homomorphic encryption lets you perform computations on encrypted data without ever decrypting it. Imagine a locked safe where someone can rearrange the contents without opening the lid. H33-128 uses BFV lattice-based FHE to compare biometric templates, run inner products, and verify identities while the data stays encrypted the entire time. The server never sees plaintext.
How do 32 users fit inside a single ciphertext?
BFV supports SIMD (Single Instruction, Multiple Data) batching. With polynomial degree N=4096 and 128-dimensional biometric vectors, we pack 4096 / 128 = 32 user templates into a single ciphertext. One FHE inner-product operation verifies all 32 users simultaneously in ~939 microseconds, which is why per-auth latency drops to roughly 36 microseconds.
What does NIST Level 1 mean?
NIST security levels classify how hard a cryptographic scheme is to break. Level 1 is equivalent to the difficulty of brute-forcing AES-128. H33-128 targets Level 1 with N=4096 and a 56-bit modulus, which provides 128-bit post-quantum security. This is the recommended baseline for most production workloads and the same level used by the NIST-standardized ML-KEM and ML-DSA algorithms.
What is the actual latency per authentication?
The full pipeline processes a 32-user batch in approximately 1,128 microseconds: ~939 microseconds for the FHE batch inner product, a near-zero in-process ZKP cache lookup, and ~189 microseconds for Dilithium attestation. Divided across 32 users, that works out to roughly ~35.25 microseconds per authentication. On Graviton4 with 96 workers running in parallel, this sustains over 2.21 million authentications per second.
How does H33-128 compare to Microsoft SEAL?
SEAL is an excellent research library for FHE primitives, but it only covers the encryption layer. H33-128 is a production pipeline that includes FHE batch matching, STARK zero-knowledge proof verification, and Dilithium post-quantum attestation in a single API call. In head-to-head throughput on equivalent hardware, H33-128 delivers 23.6x higher throughput than SEAL, largely due to Montgomery NTT, Harvey lazy reduction, SIMD batching, and NTT-domain persistence optimizations.
What do STARK proofs add to the pipeline?
STARK (Scalable Transparent ARgument of Knowledge) proofs provide a verifiable guarantee that the FHE computation was performed correctly. Without ZKP, a malicious server could return a fabricated match result. With STARK proof verification in the pipeline, any third party can confirm the authentication result is genuine. H33 caches these proofs in-process via DashMap for near-zero lookup latency.
Can H33-128 run on AWS, GCP, and Azure?
Yes. H33-128 is a cloud API, so you call it from any environment. The backend runs on AWS Graviton4 (ARM) for maximum throughput, but the API is cloud-agnostic. Your application on GCP, Azure, on-prem, or edge devices simply sends HTTPS requests to api.h33.ai. For dedicated deployments, we offer private instances on your preferred cloud provider.
What SDK languages are supported?
The core engine is written in Rust for maximum performance and memory safety. SDKs and client libraries are available for Rust, Python, JavaScript/TypeScript, and Go. The REST API can be called from any language that supports HTTPS. Mobile SDKs for iOS (Swift) and Android (Kotlin) are available via the ARM-native product for on-device encryption.
What are the 3 trust tiers?
H33-128 offers three API tiers with escalating cryptographic guarantees. H0 (Turbo) provides raw FHE biometric matching at 1 credit per operation for speed-first workloads. H1 (Standard) adds STARK zero-knowledge proof verification at 3 credits. H2 (Hardened) includes the full pipeline with FHE + STARK + Dilithium post-quantum attestation at 5 credits, giving you a verifiable, quantum-resistant proof chain.
What biometric template format does H33-128 accept?
H33-128 accepts 128-dimensional floating-point vectors normalized to unit length. These are standard biometric embeddings output by face recognition models (ArcFace, FaceNet), fingerprint matchers, and voice encoders. You extract the embedding client-side, encrypt it with the H33 SDK, and send the ciphertext. The server never sees the raw embedding.
How does key management work?
H33 uses threshold k-of-n decryption so no single server ever holds the complete secret key. Key shares are distributed across independent nodes. FHE public keys are used for encryption and homomorphic operations; the secret key is only reconstructed ephemerally during threshold decryption and is zeroed from memory immediately after. Key rotation is supported without re-encrypting stored templates.
Is H33-128 FIPS 140-2/3 validated?
H33-128 uses NIST-standardized post-quantum algorithms (ML-DSA for Dilithium attestation) and targets NIST Level 1 security. FIPS 140-3 validation for the full module is in progress. The underlying cryptographic primitives (SHA3-256, AES-256-GCM for transport, Dilithium for signatures) are individually FIPS-compliant. Contact us for the current certification timeline and SOC 2 audit reports.
Can I test H33-128 without paying?
Yes. The free tier includes 1,000 credits per month, which is enough for development and integration testing. Sign up at /pricing to receive your API key instantly. No credit card required. The free tier has full access to all three trust tiers so you can test the complete pipeline before committing to a paid plan.
What throughput can I expect on smaller instances?
The 2.21M auth/sec benchmark runs on a Graviton4 c8g.metal-48xl with 192 vCPUs and 96 workers. On a smaller c8g.16xlarge (32 workers), H33-128 sustains approximately 213,000 auth/sec. Throughput scales roughly linearly with core count. Even a modest 8-core instance can handle tens of thousands of authentications per second, which exceeds most production requirements.
Why is a single modulus faster than multiple moduli?
Traditional BFV implementations use CRT (Chinese Remainder Theorem) with multiple smaller moduli to handle large ciphertext spaces. H33-128 uses a single 56-bit modulus Q that fits in a native 64-bit register. This eliminates CRT reconstruction overhead, reduces NTT transforms per operation (one modulus instead of many), and keeps the entire working set in L1 cache. The tradeoff is lower multiplicative depth, but biometric inner products only need depth 1, making single-modulus the optimal choice.
TECHNICAL DEEP DIVES

Go Deeper

🔬 BFV vs CKKS
BFV vs CKKS: Choosing the Right FHE Scheme
Deep comparison of exact vs approximate arithmetic — when to use each scheme and why H33-128 runs BFV.
Read Full Article →
BENCHMARK
Batch Authentication: 8 Million Per Second
How SIMD batching, NTT-domain operations, and 32-user ciphertexts achieve record throughput.
Read Full Article →
🧮 NTT DEEP DIVE
NTT Performance: The Engine Inside H33-128
Montgomery multiplication, radix-4 butterflies, and Harvey lazy reduction — the math behind microsecond FHE.
Read Full Article →

Start Building With H33-128

BFV homomorphic encryption at 1.595 million auth/sec. Free tier included. No credit card required.

Get Free API Key → Read Documentation
Verify It Yourself