Secure every bit of web traffic through cryptographic means. Complete privacy, scalable computation, and zero-knowledge infrastructure for the next generation of the internet.
Watch the Quilibrium explanation video to learn about the next generation of private, decentralized computing.
Explore the architecture powering the decentralized internet layer
A simple explanation of the decentralized internet layer and how it works.
Learn the basics in simple terms
Join the decentralized internet layer that prioritizes privacy, scalability, and true cryptographic security.
Built from the ground up for privacy, scalability, and true decentralization. Not a blockchainโa complete reinvention.
Zero-knowledge by design. The network cannot see computations or data. Complete nescience through MPC and encrypted storage.
VDF-linked temporal ordering enables parallel execution and cryptographic proof sequencing without blocks or linear chains.
RDF semantic validation supports complex relationships and knowledge graphs impossible on traditional blockchains.
Random Permutation Matrix routing provides traffic analysis resistance beyond Tor, with complete sender-receiver unlinkability.
Shared-nothing architecture scales to petabyte storage and 200k+ messages/sec without compromising decentralization.
Anti-ASIC consensus rewards useful computation. Hardware diversity ensures broad participation from Raspberry Pi to servers.
See how Quilibrium stacks up against major blockchain platforms across key dimensions.
Comparative analysis of supply models across major blockchain networks. Quilibrium features a dynamic, demand-responsive emission schedule.
Unlike fixed schedules, Quilibrium adjusts issuance based on network difficulty and demand
Prevents "shrinking security budget" through work-proportional rewards
100% fair launch with no premine, VC allocation, or team reserves
| Feature | Bitcoin | Ethereum | Solana | ICP | Bittensor | Quilibrium |
|---|---|---|---|---|---|---|
| Supply Model | Fixed Cap | Unlimited + Burn | Disinflationary | Deflationary | Fixed 21M Cap | Dynamic/Adaptive |
| Current Supply | ~19.9M | ~120M | ~621M | ~549M | ~11M | ~1.35B |
| Max Supply | 21M | Unlimited | Unlimited | Unlimited | 21M | ~1.75B (2035) |
| Emission Schedule | Halving (4yr) | Proof of Stake | Disinflationary | Node rewards | Yuma Consensus (PoI) | Difficulty-triggered |
| Team/VC Allocation | 0% (Fair) | Premine + ICO | ~15% Foundation | Significant | Founder allocation | <1% (Fair) |
| Value Accrual | Scarcity | Burn + Staking | Network usage | Cycle burning | AI model quality | Computation demand |
Quilibrium implements cutting-edge cryptography from the ground up, including post-quantum resistant algorithms and multi-party computation environments that keep data encrypted during processing.
Forward secrecy and post-compromise security for all communications, extending Signal Protocol properties.
Compute on encrypted data without decryption. Enables private ML, confidential finance, and sealed-bid auctions.
SHA256/SHAKE128-based VDFs provide unforgeable timestamps and storage proofs resistant to parallelization.
Join the network by running a Quilibrium node. Earn rewards through Proof of Meaningful Work while contributing to a decentralized internet layer.
Download the latest node binaries for your platform
Download NowComplete guide to running and configuring your node
Read DocsMonitor network status and your node's performance
View DashboardMulti-core processor (ARM64 or x86_64)
More cores = more workers
SSD recommended for worker data
Regular backups required
Stable internet connection
Specific ports required
Sufficient RAM for workers
Scales with worker count
| Port | Protocol | Purpose |
|---|---|---|
| 8336 | QUIC/UDP or TCP | Master process P2P communication |
| 8340 | TCP | Master process streaming |
| 25000-25003 | QUIC/UDP or TCP | Worker processes P2P (per worker) |
| 32500-32503 | TCP | Worker processes streaming (per worker) |
config.yml and keys.yml files securely - they are required to access rewards
worker-store/ to avoid penalties
Dynamic, demand-responsive supply model with 100% fair launch distribution. Three emission scenarios model miner participation and network growth through 2033.
Projected QUIL supply and World State growth (2026โ2033)
Lower miner participation leads to slower emission growth. Supply rises from 1.35B to 1.57B, with the steepest climb between 2026โ2028 before plateauing as difficulty adjusts.
Baseline scenario with moderate participation. Supply reaches 1.59B by 2033. The curve follows natural network adoption with ASERTi3-2d difficulty adjustment ensuring sustainable incentives.
High participation accelerates emissions. Supply grows to 1.72B by 2033. The extended growth period reflects increased network difficulty and generational milestone unlocks driving faster token issuance.
The World State represents total data stored across the Quilibrium network. Starting at 143 GB in early 2026, it grows linearly to 1.8 Billion GB by 2033 as the network scales to handle petabyte-level storage across horizontal shards.
Unlike fixed-schedule cryptocurrencies, Quilibrium uses ASERTi3-2d difficulty adjustment with generational milestones. New emission phases unlock when network-wide difficulty reaches thresholds, ensuring sustainable incentives as technology advances.
Real-time network sharding analysis. Shards are grouped by ring level based on worker count, indicating competition intensity and reward potential.
Interactive 3D visualization of Quilibrium network shards
Active Workers Only: Shows only confirmed, participating workers. Excludes joining, leaving, and rejected states.
Shards visualized across a 3D globe with real-time positioning
Rotate, zoom, and click shards to explore network details
Real-time shard metrics from Quilibrium Explorer API
Auto-detects latest version from network | Shows current distribution
| Rank | Version | Build (Dec) | Peers | Percentage | Status | Distribution |
|---|---|---|---|---|---|---|
|
Loading version distribution...
|
||||||
End-to-end encrypted ML primitives and runtime. Private inference, sub-linear training, and secure multi-party computation for the next generation of AI.
10 specialized crates working in harmony for private, efficient machine learning
Foundation crate with LSH hash families, sparse tensors, and SLIDE network training primitives.
Hardware acceleration with SIMD vectorization, BF16 quantization, and cache-aligned memory.
Learnable hash functions with adaptive rebuild scheduling and drift detection.
LSH hyperparameter autotuning and sparse inference optimizations.
Transformer sparsity prediction for attention heads and MLP neurons.
LLaMA-compatible LLM inference engine with GQA, RoPE, RMSNorm, and SwiGLU.
Distributed Point Functions using AES-based BGI construction and DCF.
Two-party computation with fixed-point arithmetic and Beaver triples.
End-to-end private LLM inference via 2PC with Ferret OT and Ristretto255 OPRF.
Sub-LInear Deep learning Engine (SLIDE) uses Locally Sensitive Hashing (LSH) to achieve training and inference complexity that scales sub-linearly with network size.
Multiple LSH tables for high recall candidate selection
Per-table hash functions for bucket assignment
Dynamic LSH index rebuilding with drift detection
SLIDE achieves up to 3.5x speedup over TensorFlow on CPU for large output layers, with complexity O(LยทKยทC) instead of O(N) where C << N.
Secure two-party computation for confidential AI inference
~4.6 KB/token
~2 MB/token, ~34K triples
Production-ready LLaMA-compatible inference with optional sparsity
GQA, RoPE, RMSNorm, SwiGLU - full LLaMA compatibility
50% head sparsity, 50% neuron sparsity with Deja Vu prediction
Works with any LLaMA-architecture model in safetensors format
| Parameter | Default | Description |
|---|---|---|
| temperature | 0.7 | Sampling temperature (0.0 = greedy) |
| top_k | 40 | Top-k filtering (0 = disabled) |
| top_p | 0.9 | Nucleus sampling (1.0 = disabled) |
| repetition_penalty | 1.1 | Penalize repeated tokens (1.0 = disabled) |
| max_new_tokens | 512 | Maximum tokens to generate |
| template | auto | Chat template (zephyr, chatml, llama2, llama3, mistral, raw) |
Criterion.rs performance analysis of klearu-core primitives
| Benchmark | Median Time | Throughput | Rating |
|---|---|---|---|
| Sparse dot product (dim=1024, 10% density) | ~68 ns | ~14.7M ops/sec | Excellent |
| Sparse โ Dense conversion (dim=1024) | ~134 ns | ~7.5M ops/sec | Excellent |
| Dense โ Sparse conversion (dim=1024) | ~1.5 ยตs | ~667K ops/sec | Good |
| SimHash (dim=128, 4 tables) | ~6.9 ยตs | ~145K hashes/sec | Good |
| SimHash (dim=1024, 4 tables) | ~68 ยตs | ~14.7K hashes/sec | Expected |
| SLIDE forward pass (128โ64โ10) | ~12.6 ยตs | ~79K passes/sec | Excellent |
| SLIDE train step (128โ64โ10) | ~19.5 ยตs | ~51K steps/sec | Excellent |
| LSH query union (1000 neurons) | ~128 ยตs | ~7.8K queries/sec | Acceptable |
~68 ns sparse dot product at 1024 dimensions with 10% density (~102 nonzero values). That's roughly 1.5 ns per multiply-add โ approaching peak SIMD throughput.
This is what enables sub-linear inference speed in SLIDE. The cache-aligned memory layouts
in ContiguousWeightStore clearly pay off.
LSH query at ~128 ยตs for 1000 neurons is the main bottleneck. At production scale with 100K+ neurons per layer, this dominates.
Use klearu-bolt autotuner to reduce K (hashes)
and L (tables) to hit target recall while minimizing query cost.
Scales linearly with dimension as expected (O(dim ร K ร L)). 8ร dimension = ~10ร time.
Recommendation: For high-dimensional dense vectors, prefer SRP (Sparse Random Projection) or DwtaHash for better throughput โ they skip zero entries.
~79K passes/sec
~51K steps/sec (+54% overhead)
Train step is only ~54% slower than forward pass โ the backward pass + Adam/SGD weight update adds very little overhead. Gradient computation stays within the sparse activated set.
For a pure-CPU sparse deep learning engine, these numbers are solid and competitive.
The core primitives (sparse dot product, forward pass) are well-optimized. The LSH query latency is the area
to watch at scale โ use klearu-bolt autotuning to tune K/L and keep it in check.
This is not trying to beat CUDA GEMM for dense workloads; it's targeting the sparse CPU inference niche,
and within that niche it delivers.
Build and run Klearu in minutes
git clone https://github.com/QuilibriumNetwork/klearu.git
cargo build --release
cargo build --release -p klearu --features full
huggingface-cli download HuggingFaceTB/SmolLM-135M-Instruct
cargo run --release --bin chat -- ./SmolLM-135M-Instruct
cargo run --release --bin diagnose -- ./SmolLM-135M-Instruct