Hey Hugging Face community!
We’re excited to officially introduce ZY-S2, a lightweight, open-source AI model developed by ZYR AI, focused on coding performance, efficiency, and accessibility.
Our goal is simple: Small models deserve big capabilities.
Model: zyr-AGENT/ZY-S2 · Hugging Face
ZY-S2 is a 1.5-billion-parameter coding model built through fine-tuning Qwen2.5-Coder-1.5B.
Rather than focusing on increasing parameter counts, we explored how far targeted post-training could push a compact model.
Key highlights:
1.5B parameters
Optimized primarily for Python coding
Trained with a focus on code generation and problem-solving
Supports local CPU inference
Available in GGUF and Safetensors formats
Open-source under Apache 2.0
We evaluated ZY-S2 against established small language models using our SLMFrontier coding benchmark.
| Model | Coding Score | HumanEval+ | MBPP+ |
|---|---|---|---|
| ZY-S2 (1.5B) | 68.0% | 74.2% | 63.7% |
| Qwen2.5-Coder-1.5B | 63.5% | 63.2% | 63.7% |
| Qwen2.5-Coder-0.5B | 47.0% | 52.1% | 43.5% |
| Llama-3.2-1B | 38.0% | 27.0% | 45.6% |
All models in this coding comparison were evaluated using the same CPU llama.cpp harness and Q5_K_M quantization.
ZY-S2 achieved a 68.0% coding score, compared with 63.5% for its 1.5B Qwen baseline.
These are promising results for a compact model, although confidence intervals overlap and performance on other tasks varies.
You can run the quantized model using Ollama:
ollama run hf.co/zyr-AGENT/ZY-S2:Q5_K_M
The Q5_K_M GGUF is approximately 1.1 GB, making local experimentation possible without expensive GPU hardware.
ZY-S2 is not a general-purpose frontier model. Its main strength is Python coding, while mathematical reasoning, tool use, agentic tasks, and safety remain areas for improvement.
We believe transparent evaluations, including limitations, are essential to meaningful open-source AI development.
We’re sharing ZY-S2 with the community to encourage independent testing, feedback, and collaboration.
We’d love to hear:
How does ZY-S2 perform on your coding tasks?
What should we prioritize in future releases?
Which benchmarks would you like to see next?
What would make compact coding models more useful for real-world developers?
If you find our work interesting, consider supporting the project by liking the model repository or following ZYR AI on Hugging Face. Your feedback and support mean a lot!
Try ZY-S2: zyr-AGENT/ZY-S2 · Hugging Face
Built by ZYR AI — Pushing the boundaries of small language models.