Write a function
Describe it in a sentence and watch it stream.
> write is_even(n)
def is_even(n):
return n % 2 == 01.18B parameters✦
PrefixLM✦
4-bit GGUF✦
≈15.9 tok/s on 8 threads✦
CPU only✦
Apache-2.0✦
Thinking panel✦
Off · Low · High effort✦
OpenAI-compatible server✦
Built in India✦
→Why ZY-S1
ZY-S1 reads your whole prompt at once, thinks in a short plan, then streams the answer on the processor you already own.
The product
It plans, then it writes.
On Low or High effort a short numbered plan streams into the Thinking panel before the answer.
zyr3.vercel.app/chat
Write a function is_even(n) that returns True if n is even.
ZY-S1thinking · low
- 1Take n as input
- 2Return n % 2 == 0
- 3Check: handles negatives
def is_even(n):
return n % 2 == 0Reply to ZY-S1…
Capabilities
Small model. Sharp focus.
Everything ZY-S1 does, it does in the open and on hardware you already own.
01 / 06
Small functions, written fast
Ask for one self-contained function in Python, JavaScript, TypeScript, C++, Go or Rust. It answers in a highlighted block you can copy.
def is_even(n):
return n % 2 == 002 / 06
A plan before the code
On Low or High, a short numbered plan streams into the Thinking panel first. Read it, then collapse it.
1. Take n as input 2. Return n % 2 == 0 Check: handles negatives
03 / 06
Off, Low or High effort
Pick how much thinking you want per message. Off answers straight away, High adds a check round.
off → 512 tok low → 768 tok high → 1,024 tok
04 / 06
Streams as it types
Every token streams as it is generated, so you watch the answer build instead of waiting for it.
def square(x):
return x * x▍05 / 06
Routed to the right tool
The gateway sends code to ZY-S1, arithmetic to a calculator so numbers come back exact, and everything else to a 0.5B helper.
write is_even(n) → ZY-S1 15% of 240 → Calculator
06 / 06
No GPU required
The 4-bit build is about 700 MB and needs about 1.3 GB of RAM. An ordinary processor is enough.
zy-engine · CPU only -t 8 threads
→How ZY-S1 thinks
From your message to the answer.
- 01
Prompt
You ask for something. The gateway routes it to ZY-S1, the calculator or the helper.
- 02
Think
On Low or High, a short numbered plan streams live into the Thinking panel. High adds a check round.
- 03
Answer
ZY-S1 reads the whole prompt bidirectionally, then streams the answer left to right.
Effort
You choose how hard it thinks.
The effort pill sits next to the input. Change it per message.
Off512 tok cap
Answers straight away. Best for quick, obvious requests.
Low768 tok cap
A short numbered plan streams into the Thinking panel, then the answer.
High1,024 tok cap
The plan plus a check round before it answers.
Runs on your CPU
15.9tok/s
About 15.9 tokens per second on 8 CPU threads with zy-engine. No GPU, no account, no network once it is downloaded.
≈700 MB
Q4_K_M file
≈1.3 GB
RAM at 2,048 ctx
~/ZY-S1-GGUF · bash
$ git clone https://huggingface.co/zyr-AGENT/ZY-S1-GGUF$ cd ZY-S1-GGUF/zy-engine$ cmake -S . -B build && cmake --build build -j --target zy-engine$ ./build/zy-engine -m ../ZY-S1-Q4_K_M.gguf -t 8 -i# CPU only · 8 threads · interactive chat> write a function add(a, b)def add(a, b): return a + b1.18Bparameters
701MiB · Q4_K_M file
4,096token context
15.9tok/s on 8 CPU threads
Run it
Five commands to local.
Build zy-engine, chat in the terminal, start an OpenAI-compatible server, or call it from Python.
# needs CMake 3.18+ and a C++17 compiler, no GPU$ git clone https://huggingface.co/zyr-AGENT/ZY-S1-GGUF$ cd ZY-S1-GGUF/zy-engine$ cmake -S . -B build && cmake --build build -j --target zy-engine$ ./build/zy-engine -m ../ZY-S1-Q4_K_M.gguf -p "Write a function add(a, b)."↔Use cases
What people ask it for.
Private by default
Your code stays yours.
Run the weights on your own machine and nothing leaves it. On this site, conversation text isn't stored; the server keeps only basic request logs to enforce rate limits.
- Chats not storedOnly time, token counts and IP address, for rate limits.
- Offline once downloadedNo GPU, no account, no network.
- Your own serverzy-engine speaks the OpenAI chat API on localhost.
| Parameters | 1.18B |
|---|---|
| Architecture | Prefix language model (PrefixLM), based on Sapient AI HRM-Text-1B |
| Attention | Prompt read bidirectionally, response written causally |
| Context | Up to 4,096 tokens |
| Weights | 4-bit GGUF · Q4_K_M ≈ 700 MB (701 MiB) |
| Memory | About 1.3 GB RAM at a 2,048-token context |
| Runtime | zy-engine, CPU only |
| Speed | ≈ 15.9 tok/s on 8 CPU threads |
| Training | Supervised fine-tuning, 1,000 steps · corrective fine-tuning, 120 steps |
| License | Apache-2.0 (declared in the repository) |
| Status | Preview, for research and evaluation |
Latest from the blog
Introducing ZY-S1
How a 1.18B prefix language model was fine-tuned for small coding tasks, and why it runs on a CPU.
FAQ
Good questions.
What is ZY-S1 good at?
Short, self-contained tasks: a single small function, or a two-sentence explanation of a concept. Longer, multi-part requests are harder for it, so ask for one thing at a time.
Why does it run on a CPU?
The 4-bit file is 701 MiB and needs about 1.3 GB of RAM at a 2,048-token context. Generation is mostly limited by memory bandwidth, so an ordinary processor is enough.
Why do llama.cpp, Ollama and LM Studio give worse answers?
ZY-S1 is a prefix language model. Stock llama.cpp runs it with causal attention only, so the prompt is read the wrong way, and apps built on it share that limitation. zy-engine reads the prompt in one bidirectional pass.
Can I use it in my own project?
The repository declares Apache-2.0. The model builds on HRM-Text-1B and a public dataset, so check their terms too before you redistribute it.
Do you keep my chats?
No. Conversation text isn't stored. The server keeps basic request logs (time, token counts, IP address) to enforce rate limits.
Why is the demo sometimes busy?
It generates one reply at a time on a single machine. Other requests wait in a short queue, and each visitor can send 8 messages a minute.
Where are the weights?
Transformers weights are at zyr-AGENT/ZY-S1. GGUF files and zy-engine are at zyr-AGENT/ZY-S1-GGUF.