---
tags:
- code
- code-generation
- code-completion
- code-llm
- python
- programming
- developer-tools
- instruct
- finetuned
- qwen
- small-model
- edge-device
- local-ai
- open-source
- humaneval
- function-calling
- cli-tool
- arche-code
- text-generation
- causal-lm
- transformers
- lightweight
- cpu-friendly
- apple-silicon
- inference
- autocomplete
- llm
- 500m
- python-code
license: apache-2.0
language:
- en
- code
spaces:
- opensynapselabs/arche-codium-playground
base_model:
- Qwen/Qwen2.5-0.5B-Instruct
pipeline_tag: text-generation
library_name: transformers
---
# Arche-Codium-500M

Compact, instruction-finetuned code generation model built on **Qwen2.5-Coder-0.5B-Instruct**. Fast local code completion with minimal resources.

## TL;DR

- **500M parameters** — runs on CPU, MPS, or low-VRAM GPU
- **80% pass rate** on HumanEval (16/20 tasks)
- **Apache 2.0** — fully open, commercially usable
- **CLI-ready** — plug into [arche-code](https://github.com/OpenSynapseLabs/arche-code)

## Live Demo

Try the model directly in your browser — no setup required:

[![Open in Spaces](https://huggingface.co/datasets/huggingface/badges/resolve/main/open-in-hf-spaces-md.svg)](https://huggingface.co/spaces/opensynapselabs/arche-codium-playground)

## Quick Start

### With arche-code CLI

```bash
git clone https://github.com/OpenSynapseLabs/arche-code.git
cd arche-code
pip install -e .
arche --provider arche --model arche-codium-500m write "def fibonacci(n):" --max-tokens 256
```

The CLI auto-downloads the model on first use. Full docs: [github.com/OpenSynapseLabs/arche-code](https://github.com/OpenSynapseLabs/arche-code)

### With transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "OpenSynapseLabs/arche-codium-500m",
    torch_dtype="auto", device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("OpenSynapseLabs/arche-codium-500m")

prompt = '''def has_close_elements(numbers: list[float], threshold: float) -> bool:
    """Check if any two numbers are closer than threshold."""'''

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.2)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Benchmarks

| Benchmark | Result |
|-----------|--------|
| **HumanEval** | **16/20 (80%)** |

## What This Model Is

- **Lead magnet** — free, capable entry point into the Arche ecosystem
- **Edge-friendly** — runs on laptops, Raspberry Pi, mobile devices
- **Real code** — generates executable Python, not just snippets

## What This Model Is Not

- A replacement for 7B+ models on complex architecture tasks
- A chat model — instruction-tuned for code generation only
- The final word — larger Arche coding models are shipping this month

## Model Details

| Property | Value |
|----------|-------|
| Base model | Qwen2.5-Coder-0.5B-Instruct |
| Parameters | 0.49B |
| License | Apache 2.0 |
| Training | Instruction fine-tuning on code-completion tasks |

## Hardware Requirements

| Device | VRAM/RAM | Speed |
|--------|----------|-------|
| Apple Silicon (MPS) | 2 GB unified | ~50 tok/s |
| NVIDIA GPU (CUDA) | 2 GB | ~80 tok/s |
| CPU only | 4 GB RAM | ~10 tok/s |

## Limitations

- Struggles with multi-step reasoning (e.g., LRU cache with TTL)
- May truncate output at `max_tokens` limits — increase if code cuts off
- Hallucinates imports occasionally — always verify generated code
- Best for functions under 50 lines; breaks down on large classes

## Citation

```bibtex
@software{arche_codium_500m,
  author = {Open Synapse Labs},
  title = {Arche-Codium-500M: Compact Code Generation Model},
  year = {2026},
  url = {https://huggingface.co/OpenSynapseLabs/arche-codium-500m}
}
```

## Contact

📧 opensynapselabs@proton.me  
🐙 [github.com/OpenSynapseLabs](https://github.com/OpenSynapseLabs)

---

*Built by Open Synapse Labs. Base model: Qwen2.5-Coder-0.5B-Instruct (Apache 2.0).*