---
base_model: vsan/tiny-pickle-v3-coder
library_name: mlx
pipeline_tag: text-generation
license: apache-2.0
tags:
- mlx
- qwen3-coder
- coding
- software-engineering
- quantized
- 4bit
- apple-silicon
- moe
- tiny-pickle
---

# Tiny Pickle v3 Coder — MLX 4-bit

Tiny Pickle v3 Coder is a coding-focused adaptation of
`Qwen/Qwen3-Coder-30B-A3B-Instruct`, converted to MLX and quantized for
Apple Silicon.

## Model lineage

- Base model: `Qwen/Qwen3-Coder-30B-A3B-Instruct`
- LoRA adapter: `vsan/tiny-pickle-v3-coder-LoRA`
- Merged Safetensors: `vsan/tiny-pickle-v3-coder`
- Quantization: MLX affine 4-bit
- Group size: 64
- Converted directory size:  16G

## Installation

```bash
pip install -U mlx-lm
```

## Interactive chat

```bash
mlx_lm.chat --model mlx-community/Tiny-Pickle-v3-Coder-4bit
```

## Command-line generation

```bash
mlx_lm.generate \
  --model mlx-community/Tiny-Pickle-v3-Coder-4bit \
  --prompt "Write a tested Python implementation of an LRU cache." \
  --max-tokens 800
```

## Intended use

Code generation, debugging, code review, implementation planning, test
generation, and local software-engineering assistance on Apple Silicon.

## Quantization

This release uses MLX 4-bit affine quantization with group size 64.
Quantization reduces storage and unified-memory requirements but may alter
outputs or reduce quality relative to the merged BF16 model.

## Limitations

Tiny Pickle v3 Coder is experimental and has not yet been independently
demonstrated to outperform its base model. Generated code may be incorrect,
insecure, incomplete, or non-functional and must be reviewed and tested.

## Related repositories

- LoRA: https://huggingface.co/vsan/tiny-pickle-v3-coder-LoRA
- Merged Safetensors: https://huggingface.co/vsan/tiny-pickle-v3-coder
- GGUF: https://huggingface.co/vsan/tiny-pickle-v3-coder-GGUF

## Local Performance

The following result is a single local inference measurement, not a standardized benchmark.

| Property | Result |
|---|---:|
| Hardware | Apple M1 Max |
| Unified memory | 64 GB |
| Model format | MLX 4-bit affine |
| Quantization group size | 64 |
| Prompt length | 118 tokens |
| Prompt processing speed | 119.328 tokens/s |
| Generated length | 748 tokens |
| Generation speed | 63.481 tokens/s |
| Peak unified memory | 17.393 GB |

### Test prompt

> You are reviewing a Python async web crawler. Implement a complete, production-quality crawler that uses asyncio and aiohttp; limits global concurrency to 20; limits each domain to 2 concurrent requests; respects robots.txt; retries HTTP 429 and 5xx responses with exponential backoff and jitter; avoids duplicate URLs; normalizes relative links; restricts crawling to the starting domain; supports cancellation; records failures without stopping the crawl; and includes pytest tests using mocked HTTP responses. Return one self-contained Python module followed by the tests.

Results can vary with the macOS version, MLX-LM version, background processes, context length, sampling configuration, and thermal state.