Stop paying cloud GPUs for simple decisions.
13.4 µs latency. 74.89% accuracy (#1 Worldwide). $0.00 bills.

The Four Execution Paradigms Explained

Zev dynamically balances speed, hardware utilization, and reasoning depth. ~88% of standard traffic settles in microseconds on pure SIMD, escalating ambiguous edge cases to Apple Silicon ANE or Gemma 4.

Zev vs. Industry Competitors & Benchmark Heaven

Evaluated dynamically on the standardized frozen 231-decision JevBench suite (datasets/jevbench_public/). Comparing accuracy, median latency, throughput, cost, and hardware requirements:

Task Family Breakdown: Zev vs. Laya vs. Kev

Detailed accuracy breakdown across all 18 task families in the JevBench public evaluation suite. Headers scroll with you as you navigate through the table:

Zev Synthetic Scale Suite: High-Throughput Lexical Benchmark (1,200 Tasks & Held-Out Test Split)

Comprehensive empirical evaluation of all 7 execution paradigms across the 1,200 synthetic scale tasks (datasets/zev_benchmarks/zev_benchmarks.jsonl) and the 324 held-out test split (datasets/zev_benchmarks/zev_test_split.jsonl) on Apple Silicon, stress-testing zero-token keyword extraction throughput and memory stability under load:

Part 1: Synthetic Scale Suite (1,200 Evaluation Tasks)

Part 2: Held-Out Test Split (324 Tasks)

Adversarial Edge-Case Stress Suite (14 Exhaustive Scenarios)

Evaluated against complex prompt injection, sarcasm, clinical differentials, and state reversals:

Architectural Principles

Engineered for extreme reliability, strict guardrails, and deterministic outcomes in high-throughput production environments.

Four Ways to Run Zev

Run Zev via npx instant CLI, Homebrew binary, HTTP REST daemon, or embedded Rust library across all four execution modes.

# Method 1: Interactive First-Run Onboarding & Hardware Diagnostics
zev init          # Dynamic hardware prober & interactive method wizard
zev doctor        # Inspect SIMD, Apple Neural Engine, and Gemma endpoints
zev config show   # View persistent ~/.config/zev/config.json

# Method 2: Instant CLI Routing (Microsecond Latency)
brew install zev  # Or: cargo install zev-rs
zev route \
  --state "Critical: production database is down with 500 connection refused errors" \
  --routes '{"billing": "Invoices", "infra": "Database outages", "sales": "Enterprise"}'

# Method 3: High-Performance REST API Server (Axum / Standalone Daemon)
zev serve --port 8080
# Drop-in HTTP API on http://127.0.0.1:8080/v1/decisions

# Method 4: Streaming Batch Pipelines (High-Throughput JSONL)
cat tickets.jsonl | zev batch --file - --mode route \
  --routes '{"billing": "Billing", "infra": "Infra", "sales": "Sales"}' | jq

// Add to Cargo.toml:
[dependencies]
zev-rs = { version = "0.3.9", features = ["neural"] }

use zev::{ZevEngine, ApfelNeuralBackend, ExecutionMode};

let engine = ZevEngine::default();
let backend = ApfelNeuralBackend::new();

// MODE 1: Load-Balanced Production Reflex (74,935 ops/s • 13.3 µs • Automatic Failover)
// Alternates sub-threshold traffic across ANE and Gemma 4 with automatic bidirectional failover
let response = engine.evaluate_recommended(&request, &backend)?;

// MODE 2: Bayesian Product of Experts Ensemble (74,346 ops/s • 13.4 µs • 74.89% Accuracy — #1 Worldwide)
// Simultaneous dual neural execution with Bayesian PoE consensus; eclipses Winnow-12B #1
let response = engine.evaluate_recommended_ensemble(&request, &backend)?;

// MODE 3: Three-Tier Sequential Cascade (73,806 ops/s • 13.5 µs • < 16 MB RSS Memory)
// Sequential escalation: SIMD reflex (>0.85) → Apple ANE (>0.75) → Gemma 4
let mode = ExecutionMode::Cascade { fast_threshold: 0.85, neural_threshold: 0.75 };
let response = engine.evaluate_with_mode(&request, mode, &backend)?;

// MODE 4: Pure Hardware SIMD Reflex (75,078 ops/s • 13.3 µs • Zero Model Weights)
// Pure CPU Neon/AVX2; zero weights, zero PyTorch/GPU. Runs on any Docker container or VPS
let response = engine.evaluate(&request)?;

// Or instant 5.8 µs zero-allocation routing:
let (route, confidence) = engine.route("Refund request #1084", [
    ("billing".into(), "Refunds and invoices".into()),
    ("tech_support".into(), "API errors and bugs".into()),
])?;