You describe it.
Your AI writes Zyth.
Zyth was designed for models to write. Every error says exactly what to change, the tools apply the change, and the whole language reference fits in one prompt.
Tasks right, first try / after fixes
Bars: first try. Temperature 0, one run each, up to three rounds of checker, run and test feedback. Measured 2026-10-29.
What the numbers say
The two strongest models write Zyth as well as Python, first try and after fixes. The others trail by 5 to 15 points on the first try and are level or close once they see what the checker says.
A third of the tasks were held out: never looked at while Zyth was being changed, so the result is not tuned to the test. The tasks cover text, lists, maps, parsing, graphs, tensors and data tables.
Models are not new to Zyth's habits: common spellings from Rust, Python and JavaScript (vec![..], x in xs, and/or, .iter().collect()) mean the same thing, and zyth fmt rewrites them to Zyth's own.
Errors with the fix
zyth check --json gives each error with the exact edit that fixes it. zyth fix applies them.
"code": "E0406",
"fixes": [{ "line": 3, "col": 18,
"replace": "256", "with": "512" }]An MCP server
zyth mcp lets an AI agent check, run, test and fix Zyth code in your project through the Model Context Protocol.
zyth mcp
A reference for models
The whole language in one compact file that fits in a prompt: llms.txt.
curl https://zyth.io/llms.txt
Write a program from a sentence
zyth make writes a program, checks it and fixes it until it runs.
zyth make "a to-do app" -o todo.zyth
Convert Python
zyth convert rewrites a Python or JavaScript program in Zyth, and keeps going until it prints the same output.
zyth convert prog.py
AI decisions in code
A model picks one of the answers you wrote, within a budget, and every decision can be replayed.
decide "Spam?" from msg within(dollars: 0.001) {
spam => archive(msg)
ok => deliver(msg)
unsure => ask_a_person(msg)
}Harder tasks, measured in November
A third set of 51 tasks added projects of several files, Python and npm packages, data handling and web servers. The table counts only the 63 held-out tasks of all three sets, which were never used to change Zyth. First try / after fixes, Zyth against Python, 2026-11-13:
| Model | Zyth | Python |
|---|---|---|
| Claude Sonnet 5.5 | 98% / 100% | 100% / 100% |
| GPT-6.1 Terra | 98% / 100% | 100% / 100% |
| DeepSeek V3.2 | 89% / 94% | 92% / 100% |
| Qwen3 Coder 30B | 76% / 92% | 89% / 97% |
| Llama 3.3 70B | 67% / 90% | 89% / 94% |
The strongest models are within a task of Python. The smaller ones still trail, most on the new kinds of task, and closing that gap is the work in progress: fixes taken from the models' own mistakes on tasks that are not held out, and a sharper llms.txt, raised the first try on the held-out tasks from 140 to 147 of 189 answers for the three smaller models.
Cheaper decisions that stay right
ai.cheaper(main, cheap) starts with the strong model and moves to the cheap one where they agree. On 40 messages to sort into spam or not: Claude Sonnet alone, 40 of 40 right for $0.032; with Llama 3.3 70B taking over after agreeing on the first 8, still 40 of 40, for $0.012.
zyth run --diary run.jsonl records every decision, --replay repeats them exactly, and zyth diary suggests plain code for decisions that always come out the same.