README.md

---
license: mit
language:
- en
- ja
tags:
- gguf
- llama.cpp
- code-generation
- javascript
- typescript
- react
- nodejs
- slm
pipeline_tag: text-generation
---

# AI LocalQmod — JS-family Code SLMs (GGUF)

**Ultra-small, single-purpose code-generation models — as light as ~17 MB.**

This repository hosts a family of small language models (SLMs) trained from scratch
specifically for JavaScript-ecosystem code generation. They are built and distributed by
**AI LocalQmod** for use with [AI-App Builder](https://ai-localqmod.com), a local-first
AI app generation tool, but the GGUF files are plain [llama.cpp](https://github.com/ggerganov/llama.cpp)
models and can be loaded with any GGUF-compatible runtime (llama.cpp, llama-cpp-python,
Ollama with a Modelfile, LM Studio, etc.).

Unlike general-purpose LLMs, each model here is trained on a narrow, single-purpose corpus
(natural-language instruction → JS-family code) with a from-scratch tokenizer tuned for
that language's syntax and vocabulary. This trades general reasoning ability for
extremely small size and fast local inference, making them a good fit for lightweight,
single-purpose code-completion helpers rather than full coding assistants.

## Models

| File | Target | Quantization | Size |
|---|---|---|---|
| `js-slm-Q4_K_M-chat.gguf` | JavaScript | Q4_K_M | 17.7 MB |
| `jsts-slm-Q4_K_M-chat.gguf` | JavaScript + TypeScript | Q4_K_M | 17.5 MB |
| `node-slm-Q4_K_M-chat.gguf` | Node.js (fs/path/process/http/events) | Q4_K_M | 20.0 MB |
| `react-slm-Q4_K_M-chat.gguf` | React components | Q4_K_M | 16.6 MB |
| `js-loops-slm-Q4_K_M-chat.gguf` | JavaScript (for-loops + arrays: sum/max/count/filter/map/average) | Q4_K_M | 20.4 MB |

Sizes measured with `stat -f%z` (exact byte count), not rounded `ls -lh` output.

## Intended use

- Local, single-purpose code generation for the target language/framework shown above
- Multi-agent "orchestration" pipelines where a larger LLM plans and one of these SLMs
  generates individual small functions/components
- Environments where downloading a multi-GB model is impractical (offline demos, low-disk
  devices, quick experiments)

## Not intended for

- General-purpose chat or reasoning
- Large, multi-file refactors or architecture design
- Languages/frameworks outside each model's specific target (see table above)

## Usage (llama.cpp)

```bash
llama-cli -m js-slm-Q4_K_M-chat.gguf -p "Write a function that returns the sum of an array" -n 128
```

## Usage (llama-cpp-python)

```python
from llama_cpp import Llama

llm = Llama(model_path="js-slm-Q4_K_M-chat.gguf")
result = llm.create_chat_completion(
    messages=[{"role": "user", "content": "Write a function that returns the sum of an array"}]
)
print(result["choices"][0]["message"]["content"])
```

## License

MIT. Free to use, modify, and redistribute.

## About AI LocalQmod

AI LocalQmod builds local-first, privacy-respecting AI tools. Learn more at
[ai-localqmod.com](https://ai-localqmod.com).