FIND IT BEFORE THEY DO. · RUST GATE #![forbid(unsafe_code)] · ZERO-JS
CortexRouter · Open Collective · Compute Architecture

Scaling CortexRouter · Compute Support

CortexRouter Zero 1.0 — 119B MoE Zero-Day Discovery Model
FLAGSHIP MODEL · MISTRAL SMALL 4 (119B MoE) · GATED WEIGHTS & PRIVATE API

CortexRouter Zero 1.0 — 119B MoE Zero-Day Discovery Model

Trained on Mistral Small 4 (119B total parameters, 128-expert hybrid Mixture-of-Experts architecture, 256K context window), CortexRouter Zero 1.0 performs deep multi-step reasoning across full source repositories, ASTs, and protocol state machines to uncover authentication bypasses, cryptographic flaws, and memory-safety vulnerabilities.

Mistral Small 4 · 119B MoE 128 Experts · ~6.5B Active 256K Context Window Gated Weights · Unpublished Zero Cloud Leakage
🔒 Weights Not Publicly Released — API Access by Request: To prevent dual-use misuse, CortexRouter Zero 1.0 (119B) weights remain unpublished. Verified security teams, software vendors (PSIRT), CERT/CSIRT units, and enterprise defense partners may contact us to request private API access.
CortexRouter Judge v1 (Apache-2.0 · Open Weights)
LIVE ARTIFACT · APACHE-2.0 · HUGGING FACE

CortexRouter Judge v1 (Apache-2.0 · Open Weights)

Open-weights Qwen3-8B LoRA vulnerability-evidence verification model producing structured 9-field triage decisions (Confirmed / Patched / False Positive) with 80/80 (100%) accuracy on our published 20-CVE clean holdout benchmark.

80/80 (100%) Holdout Apache-2.0 Open Weights Qwen/Qwen3-8B LoRA Reproducible Benchmark
Commitment: Contributions never enter personal income. 100% of contributed funds are spent exclusively on dedicated laboratory GPU/compute hardware to train and scale CortexRouter offline, with every vendor invoice published publicly.
PHASE 01 · CORTEXROUTER CORE 20,500 EUR

1× NVIDIA RTX PRO 6000 Blackwell

96 GB GDDR7 ECC Zero Cloud Leakage ~15% Price Buffer
PHASE 02 · 4× TRANCHES 39,500 EUR

4× NVIDIA DGX Spark Founders Edition

512 GB Unified Memory 16 TB NVMe 9,875 EUR / Node

Accountable Fiscal Hosting (Open Collective)

Financial contributions for CortexRouter are structured on Open Collective (opencollective.com/cortexrouter), where the fiscal host holds funds, verifies official hardware invoices directly, and maintains an immutable public ledger.

Direct Hardware Contributions

Direct donation of compute hardware for CortexRouter (including RTX PRO 6000 Blackwell or DGX Spark units) is welcomed and represents the cleanest support path because no money changes hands. Contact vahap@cortexrouter.ai prior to shipping.

CortexRouter Laboratory Hardware Roadmap

Equipment CortexRouter Role Target Budget (incl. Buffer) Milestone
1× NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB GDDR7 ECC) Priority 1 — CortexRouter core offline zero-day model training & large-context source/binary auditing without external cloud APIs 17,900 – 20,500 EUR Phase 1 · Priority
1× NVIDIA DGX Spark Founders Edition (GB10 Grace Blackwell, 128 GB, 4 TB NVMe) Priority 2a — Dedicated CortexRouter local Grace Blackwell fuzzing, symbolic triage & zero-day variant node (Node 1 of 4) 8,500 – 9,875 EUR Phase 2 · Node 1/4
3× NVIDIA DGX Spark Founders Edition (QSFP Cluster Expansion — 4 Nodes / 512 GB Total) Priority 2b — Distributed 4-node CortexRouter training & continuous open-source zero-day discovery cluster (Nodes 2–4 of 4) 25,500 – 29,625 EUR Phase 2 · Nodes 2–4/4

Hardware Market Volatility & Unit-by-Unit Procurement

Enterprise GPU and AI workstation retail prices increase frequently (current retail baseline: ~17,899 EUR for RTX PRO 6000 Blackwell 96GB and ~8,499 EUR per DGX Spark Founders Edition, totaling ~51,895 EUR raw retail before host fees). To protect against ongoing market price hikes and cover the ~8% fiscal host fee, CortexRouter milestone targets include a ~15% buffer (20,500 EUR for Phase 1; 39,500 EUR across 4 DGX Spark tranches = 60,000 EUR cumulative ceiling). Each unit is ordered immediately as its individual milestone clears so funds do not lose purchasing power; every disbursement strictly matches the exact official retailer invoice, and any unused buffer rolls directly into the next hardware node.

Support links reside solely on this page (and ledger verification on Transparency) and are never included in vulnerability advisories or vendor correspondence.