523x faster than Pydantic. 31x faster than msgspec-ext. 20x faster than Zod. Same API.

One validation core. Three ecosystems. Zero compromise.

Python:     24,100,000 validations/sec (523x faster than Pydantic, 31x faster than msgspec-ext)
TypeScript: 20x faster than Zod 4 (avg), up to 50x on number formats
Zig:        Zero-cost. Comptime. No runtime.

from dhi import BaseModel, Field, EmailStr
from typing import Annotated

class User(BaseModel):
    name: Annotated[str, Field(min_length=1, max_length=100)]
    email: EmailStr
    age: Annotated[int, Field(gt=0, le=150)]

user = User(name="Alice", email="[email protected]", age=28)
user.model_dump()   # {'name': 'Alice', 'email': '[email protected]', 'age': 28}

- import { z } from 'zod';
+ import { z } from 'dhi';

const User = z.object({
  id: z.string().uuid(),
  name: z.string().min(1).max(100),
  email: z.email(),              // Zod 4 top-level shortcuts
  age: z.int().positive(),       // Zod 4 number formats
  createdAt: z.iso.datetime(),   // Zod 4 ISO namespace
});

const user = User.parse(data); // same API, 20x faster

Two backends — same API:

const dhi = @import("model");

const User = dhi.Model("User", .{
    .name = dhi.Str(.{ .min_length = 1, .max_length = 100 }),
    .email = dhi.EmailStr,
    .age = dhi.Int(i32, .{ .gt = 0, .le = 150 }),
});

const user = try User.parse(.{
    .name = "Alice",
    .email = "[email protected]",
    .age = @as(i32, 28),
});
// Validation is inlined at compile time. Zero allocations. Zero dispatch.

Same validation rules. Same error behavior. Three languages. One core.

Benchmarks auto-generated via CI. See raw data.

BaseModel layer: 546K model_validate/sec | 6.4M model_dump/sec

Note on msgspec: Plain msgspec (~5.8M/sec) does type-checked JSON decoding but lacks field-level validators (email, URL, positive int). msgspec-ext adds those 26 validators via Python dec_hook callbacks, bringing it to ~777K/sec — a fairer apples-to-apples comparison with dhi's validated parsing.

pip install dhi          # Python (wheels for macOS arm64 + Linux x86_64)
npm install dhi          # TypeScript (Node 18+ / Bun / Deno)

Pure Python fallback included — no native extension required to get started.

dhi is a Zig core with bindings for JS and Python:

src/              → Zig core: validators, SIMD, C API, WASM API (the engine)
js-bindings/      → npm package: Zod 4 drop-in replacement
python-bindings/  → PyPI package: Pydantic drop-in replacement
docs/             → Benchmark charts, shared docs
.github/          → CI, release workflows

Which package should I use?

• TypeScript/JS: npm install dhi — replaces Zod, works in browsers + edge + Node.js
• Python: pip install dhi — replaces Pydantic, 523x faster validation
• Zig: Import src/ directly — zero-cost comptime validation

dhi implements 100% of the Zod 4 API, including all new Zod 4 features:

z.email()      z.uuid()       z.url()        z.ipv4()       z.ipv6()
z.jwt()        z.nanoid()     z.ulid()       z.cuid()       z.cuid2()
z.base64()     z.e164()       z.mac()        z.cidrv4()     z.cidrv6()
z.hex()        z.hostname()   z.hash('sha256')

// ISO namespace
z.iso.datetime()  z.iso.date()  z.iso.time()  z.iso.duration()

// Number formats
z.int()     z.float()    z.float32()   z.float64()
z.int8()    z.uint8()    z.int16()     z.uint16()
z.int32()   z.uint32()   z.int64()     z.uint64()

z.stringbool()                           // "true"/"yes"/"1" → true
z.templateLiteral(['user-', z.number()]) // Template literal types
z.json()                                 // Recursive JSON schema
z.file().mime('image/png').max(5_000_000) // File validation
z.registry()                             // Schema registry
z.prettifyError(error)                   // Pretty error formatting

Full model_validate(), model_dump(), model_dump_json(), model_json_schema(), model_copy() support.

dhi is written in Zig — a systems language with compile-time code generation, no garbage collector, and direct hardware access. The same source compiles to:

Key tricks:

• Comptime models — Validation logic is generated at compile time. No vtables, no reflection, no hash lookups.
• SIMD batch validation — Process 4 values per cycle on 256-bit vectors.
• Single FFI call — Python batch validation crosses the FFI boundary once, not per-item.
• N-API over WASM — For Node.js, the native addon skips alloc/encode/dealloc per string call: 1.7–2x faster for format validators (url, ipv4, datetime).
• No allocations — The happy path never allocates. Errors are stack-returned.

┌─────────────────────────────────────────────────────┐
│            Zig Core (comptime + SIMD)                │
└──────┬────────────────┬──────────┬────────────────┬──┘
       │                │          │                │
  ┌────▼─────┐    ┌────▼────┐ ┌───▼──────┐    ┌───▼──────┐
  │  Python  │    │  WASM   │ │  N-API   │    │   Zig    │
  │  C ext   │    │  28KB   │ │  .node   │    │  Native  │
  └────┬─────┘    └────┬────┘ └───┬──────┘    └───┬──────┘
       │                │          │                │
  BaseModel        z.object()  z.object()      Model()
  Pydantic API    Zod 4 (edge) Zod 4 (Node)  comptime API

# Python — full comparison (dhi vs msgspec vs msgspec-ext vs satya vs Pydantic)
git clone https://github.com/justrach/dhi.git && cd dhi
cd python-bindings
pip install -e .
pip install msgspec msgspec-ext 'pydantic[email]' satya
python benchmarks/benchmark_vs_all.py

# TypeScript — dhi vs Zod 4
cd js-bindings && bun install && bun run benchmark-vs-zod.ts

# Zig
zig build bench -Doptimize=ReleaseFast

The Python benchmark validates 10,000 user objects with 4 field-level validators each:

Each library uses its idiomatic equivalent:

• dhi: _dhi_native.validate_batch_direct() — single FFI call, Zig SIMD validators
• msgspec: msgspec.json.Decoder — C-level JSON decode + type check (no format validators)
• msgspec-ext: msgspec.json.decode() with dec_hook — adds EmailStr, HttpUrl, PositiveInt validators on top of msgspec
• satya: model_validate_json_array_bytes() — Rust JSON parse + validation
• Pydantic v2: model_validate() per item — Rust-backed, with EmailStr + Field constraints

Timing uses time.perf_counter() over 5 runs after a warmup pass. Results are averaged. Tested on Python 3.13 and 3.14 (no significant difference — hot paths are native code).

dhi started as a question: can Zig's type system unify validation across language boundaries?

The hypothesis: Zig's comptime can generate the same validation semantics that Pydantic builds with metaclasses and Zod builds with method chains — but at compile time, with zero runtime cost, targeting any platform via its C ABI and WASM backends.

Results:

MIT

dhi — the fastest validation library for every language you use.