Benchmarked February 2026 - Single MacBook Pro

Sub-5ms authentication
at 1,000 concurrent sessions

FHE encryption, ZK proofs, biometric verification, and quantum signatures - all under load, all with zero errors. These numbers are our floor, not our ceiling. Benchmarked on a single MacBook Pro. Production cloud hardware only goes up from here.

3ms
Median Latency
8,099/s
Requests Per Second
0%
Error Rate
1,000
Concurrent Sessions

Three modes. Choose your tradeoff.

Select at enrollment time. Every mode delivers full homomorphic encryption with zero plaintext exposure.

Turbo
411us
full auth flow
  • Accuracy 90-95%
  • Polynomial Degree N = 2048
  • Security 128-bit
  • Encrypt 291us
  • Best for High volume
Standard
1.6ms
full auth flow
  • Accuracy 98-99%
  • Polynomial Degree N = 4096
  • Security 128-bit
  • Encrypt 1.15ms
  • Best for Production default
Precision
5.7ms
full auth flow
  • Accuracy 99.9%+
  • Polynomial Degree N = 8192
  • Security 128-bit
  • Encrypt 4.21ms
  • Best for Regulated

Fast at any load. On a laptop.

Single MacBook Pro (Apple M4 Max). No cloud infrastructure, no GPU offloading, no distributed cluster. Production deployments on multi-core cloud instances scale linearly from here.

Concurrent Users FHE Cycle Median FHE Cycle P95 Throughput Error Rate
200 3ms 51ms 1,688 req/s 0%
500 3ms 81ms 4,156 req/s 0%
1,000 3ms 412ms 8,099 req/s 0%
Benchmark Hardware
MacBook Pro - Apple M4 Max
Single laptop. No cloud instances. No GPU acceleration. No distributed cluster. These are the numbers you get before any infrastructure investment.
Production Deployment
This is the floor, not the ceiling
Production cloud instances (96+ cores, dedicated memory) scale linearly. Horizontal scaling across nodes multiplies throughput with no architectural changes.

Every operation, measured.

Median latency at 200 concurrent users on a single MacBook Pro. Each operation runs over fully homomorphic encrypted data.

FHE Context Init
Pre-warmed pool, zero cold start
319us
FHE Encrypt
BFV/CKKS with SIMD packing
299us
FHE Compute
Homomorphic distance calculation
422us
Biometric Verify
Encrypted template matching
85us
Usage Metering
Lock-free atomic path
87us
Auth Token
Verify + session lookup
222us
Full Authentication Cycle - Median Latency Breakdown
FHE Init
319us
Encrypt
299us
Compute
422us
Biometric
85us
Metering
87us
Auth Token
222us
Total Cycle
3ms

Why it stays fast at scale.

Architecture decisions that make a single MacBook Pro outperform distributed FHE clusters.

POOL
Pre-warmed FHE Contexts
BFV and CKKS contexts with NTT tables built at startup, shared via zero-copy references. No per-request initialization cost.
LOCK-FREE
Atomic Concurrency
Lock-free data structures for rate limiting, usage tracking, and auth metering. Write contention eliminated from the hot path.
COALESCE
Compute Deduplication
Identical FHE operations within a time window are computed once and broadcast to all waiters. Turns O(N) into O(unique).
CACHE
ZK Circuit Caching
R1CS circuits and verification keys cached after first generation. Batch verification with parallel parsing for logarithmic-time proof checks.
ARC
Zero-Copy Keys
400KB+ FHE key sets shared via reference counting instead of deep copying. Nanosecond access under any concurrency level.
SIMD
Packed Ciphertexts
Biometric vectors packed into BFV slots for parallel encrypted computation. Up to 8x throughput on batch operations.

Different league.

H33 on a single laptop delivers the full post-quantum stack in less time than competitors take for a single operation on production infrastructure.

Capability H33 Nearest Competitor Advantage
Biometric Verification 220us (Turbo) Onfido - seconds ~1000x faster
FHE Encryption 150us Zama - 100x slower 100x faster
ZK Proof Generation 267us (cached) RISC Zero - 44+ sec 165,000x faster
Quantum Signatures 99us (Dilithium) - No competitor
Full Stack Auth 220us - 5.7ms 17-35 sec combined 3,000-80,000x
Concurrent Sessions 1,000+ @ 0% errors Not published -

See it yourself.

1,000 free auths. Full cryptographic stack. No sales call.

npm install @h33/client