Zyth checks every tensor's shape, type and device before anything runs, tells you where each number came from and how to fix it, and then runs at Go and Rust speed, on GPUs from a laptop to a rack of H100s, in the browser and inside React.
Website: zyth.io
1.6 ms to start a built program (Java: 24.2). 1,276 TFLOP/s of fp8 on one H100 (PyTorch: 1,125). 100% of 202 tasks written right on the first try by Claude and GPT.
The playground runs the real Zyth in your browser through WebAssembly. Nothing to install.
Website · Playground · Demos · Specification · Reference for AI models · Standard library · What is in this repository · Licensing
Today, AI work is Python on top, C++ and CUDA underneath, and JavaScript in front. Zyth is one language for all of it, and its checker sees the whole program. A tensor's type says its shape, so a layer that does not fit the one before it is an error before the program runs, not a crash at step 4,000:
error[E0406]: `self.fc2` expects tensor<f32, [..., 256]>, found tensor<f32, [B, 512]>
--> model.zyth:6:16
|
6 | return x |> self.fc1 |> relu |> self.fc2
| ^^^^^^^^^^^^^^^^^^^^^ dimension 1 is 512, expected 256
|
= note: 256 comes from `Linear(256 -> 10)` at model.zyth:3
= note: 512 comes from `Linear(784 -> 512)` at model.zyth:2
= help: change line 3 to `Linear(512 -> 10)`
The same fix comes as JSON from zyth check --json, so an editor or an AI agent applies it for you.
model Net {
fc1 = Linear(784 -> 512)
fc2 = Linear(512 -> 10)
fn forward(self, x: tensor<f32, [B, 784]>) -> tensor<f32, [B, 10]> {
return x |> self.fc1 |> relu |> self.fc2
}
}
let net = Net().to(gpu)
print(net(tensor.randn(32, 784).to(gpu)).shape) // [32, 10]
| Train and fine-tune | Automatic gradients, zyth.nn, AdamW, mixed precision; LoRA, SFT, DPO and GRPO; models from the Hugging Face Hub; GGUF and ONNX in and out. |
| GPU kernels in Zyth | kernel fn compiles to CUDA for NVIDIA, HIP for AMD, and WGSL for Apple, Intel and the browser, down to tensor-core tiles. |
| Many GPUs, many machines | Data, tensor, pipeline and expert parallelism, sharded weights, fp8 training, and zyth launch across machines with restarts from checkpoints. |
| One file to ship | zyth build --release compiles to machine code through LLVM and packs the program and its runtime into one executable. |
| In the browser and React | Build a program into a web page, or a component into an npm package; models run in WebAssembly, on the GPU through WebGPU. |
| AI decisions as code | decide lets a model pick one of the answers you wrote, within a budget, with a diary you can replay exactly. |
| Packages with permissions | Zyth, Python and npm packages, each limited to what its allow list grants; every version's hash in a public log. |
| Written by AI | Errors with exact fixes, an MCP server (zyth mcp), and a reference that fits in a prompt (llms.txt). |
| Feature | Python | Mojo | Rust | Go | TypeScript | Julia | Zyth |
|---|---|---|---|---|---|---|---|
| Easy to write | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ |
| Script mode (no main) | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ |
| Compiles to machine code (1) | ❌ | ✅ | ✅ | ✅ | ❌ | JIT | ✅ |
| Fast general code (text, lists, maps) (2) | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
| Tensors as a built-in type | libraries | ✅ | libraries | libraries | libraries | ✅ | ✅ |
| Tensor shapes checked before running (3) | ❌ | Partial | ❌ | ❌ | ❌ | ❌ | ✅ |
| Automatic gradients built in | libraries | ❌ | libraries | libraries | libraries | libraries | ✅ |
| Fine-tuning built in (LoRA, SFT, DPO, GRPO) | libraries | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
| GPU kernels in the language (4) | libraries | ✅ | libraries | ❌ | ❌ | ✅ | ✅ |
| NVIDIA GPUs | libraries | ✅ | libraries | ❌ | ❌ | ✅ | ✅ |
| Apple GPUs | libraries | ✅ | libraries | ❌ | ❌ | libraries | ✅ |
| AMD GPUs (5) | libraries | ✅ | libraries | ❌ | ❌ | libraries | Written |
| TPUs (6) | libraries | Announced | ❌ | ❌ | ❌ | ❌ | Export |
| Training across machines, with restarts (7) | libraries | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
| Memory safety without a garbage collector (8) | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
| Concurrency (9) | Partial | Partial | ✅ | ✅ | ✅ | ✅ | ✅ |
| Runs in the browser (10) | libraries | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ |
| AI models on the browser's GPU (WebGPU) | ❌ | ❌ | libraries | ❌ | libraries | ❌ | ✅ |
| One file to ship (11) | ❌ | Partial | ✅ | ✅ | ❌ | Partial | ✅ |
| Packages limited by permissions (12) | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
| Runs on Windows (13) | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ |
zyth build --releasecompiles through LLVM (or Cranelift) into one executable.- Seven programs, best of 15 runs on one core, 2026-11-01: Zyth faster than Go on all seven and level with Rust on five; Rust is faster on text and sort.
- Mojo's compile-time layouts serve kernels (
LayoutTensor); it does not check a whole program's shapes. - Python's are Triton (a library with its own subset); Rust's are rust-gpu or CUDA C strings.
- Zyth's HIP backend is written but has never run on an AMD GPU.
- Zyth exports models to StableHLO (run in XLA and IREE on the CPU); never run on a TPU. Mojo: announced by Modular.
zyth launch, checkpoints and restarts.- Python, Go, TypeScript and Julia are memory safe through a garbage collector.
- Python: the GIL, unless the free-threaded build is used.
- Python through Pyodide; Rust and Go compile to WebAssembly.
mojo buildlinks Mojo's runtime libraries dynamically; Julia needs PackageCompiler.- Each package may use only the network, files and effects its
allowlist grants. - Mojo: Windows is not on its roadmap; WSL works.
"libraries" means the language can do it through packages you add. Other languages checked against their own pages on 2026-10-31. Built in, of 21: Python 3, Mojo 10, Rust 7, Go 7, TypeScript 6, Julia 8, Zyth 19. The same table is on zyth.io.
Six programs in each language, built for release, best of 15 runs on one core (Linux, 2026-11-01), ms:
| Program | Zyth | Rust | Go | Java |
|---|---|---|---|---|
| start and exit | 1.6 | 1.7 | 1.9 | 24.2 |
| recursive fib(30) | 3.2 | 3.4 | 5.3 | 27.4 |
| count 300,000 words | 21.1 | 21.2 | 24.4 | 70.0 |
| sieve, 2M elements | 5.7 | 5.8 | 7.4 | 40.3 |
| split and parse text | 25.3 | 18.7 | 37.8 | 126.0 |
| update a list of structs | 1.8 | 1.9 | 3.6 | 55.2 |
| sort 300,000 numbers | 18.7 | 10.7 | 71.9 | 110.5 |
On an NVIDIA H100 against PyTorch on the same machine: fp8 products 1,276 TFLOP/s (PyTorch 1,125); an fp8 training step 11.9 ms (torchao float8 compiled 12.7); attention forward and backward 0.66 ms (FlashAttention-2 0.91). Every number, with what is still behind: zyth.io/speed and zyth.io/gpu.
Zyth is free to download and use. The parts you build on are open source here, under Apache-2.0 or MIT, at your option:
| Part | Where |
|---|---|
| The language specification | ZYTH-SPEC.md |
| The language reference for AI models | docs/llms.txt |
| The standard library written in Zyth | stdlib/ |
| Type definitions for Python and npm packages | registry/types/ |
| The package tools: checksum log, auditor and witness, registry, package hashes | crates/zyth-pkg/ |
| Editor support (VS Code) | editors/ |
| Examples, samples and showcases | examples/, samples/, showcases/, packages/ |
The compiler, checker, interpreter and runtimes are proprietary to BioDream LLC and free to use,
including for commercial work; the programs you make with them are yours. See LICENSING.md.
Where these files mention paths such as tests/, scripts/ or crates/zyth-cli, those are in the
compiler's sources, which are not published.
The package tools build on their own, so anyone can run, check and audit the checksum log Zyth relies on:
cargo test --releasePreview. The language and its tools are still changing. The first public release of the zyth
binaries is on its way; until then, everything runs in the playground and the
demos. Then it installs with one line:
curl -fsSL https://zyth.io/install.sh | sh # Linux and macOS
irm https://zyth.io/install.ps1 | iex # Windows (PowerShell)Issues and pull requests are welcome for the parts in this repository: the specification, the standard library, type definitions, the package tools, editor support and the examples. Contributions are accepted under Apache-2.0 or MIT. Bugs in the compiler itself: open an issue with the smallest program that shows it.
BioDream LLC. BioDream builds software for work where mistakes are expensive: healthcare integration (Perfuse), phone systems and 911 (BitVoice, bitdispatcher), and rules that explain themselves (quorum, Vaglio). Zyth is the language for the AI underneath.
Questions and commercial support: firstmind@biodream.ai
zyth.io · biodream.io · github.com/BioDream-LLC
© 2026 BioDream LLC