# JANGQ — Smaller Models, Bigger MLX Wins

> Public AI-readable summary for https://jangq.ai.
> The benchmark page is intentionally strict: it only lists cases where JANG is smaller than the closest MLX baseline and wins by a large quality margin.
> Open source, Apache 2.0. GitHub: https://github.com/jjang-ai/jangq. Models: https://huggingface.co/JANGQ-AI.

## Listing Policy

JANGQ does not list every tested model on the public homepage.

Included results must satisfy all of the following:

1. JANG uses less memory, less disk, or a lower average bit width than the closest MLX comparison.
2. JANG wins by a decisive MMLU margin, coherency margin, or measured error margin.
3. The comparison uses the same base model family and a nearby MLX quantization target.
4. The result is a proven benchmark or prompt-level failure comparison, not a speculative claim.

Close wins, same-size-only wins, larger JANG profiles, and broad recent uploads are excluded from the visible model list.

## Proven Smaller-Win Set

### MiniMax-M2.5 — JANG_2L vs MLX 4-bit

The largest public win on the page.

| Method | Avg bits | GPU memory | Disk | MMLU |
| --- | ---: | ---: | ---: | ---: |
| JANG_2L | 2.10 | 82.5 GB | 89 GB | 74.0% |
| MLX 4-bit | 4.00 | 119.8 GB | 120 GB | 26.5% |

Result: JANG is 37.3 GB smaller in GPU memory and wins by +47.5 MMLU points.

Curated release: https://huggingface.co/JANGQ-AI/MiniMax-M2.5-JANG_2L

### Qwen3.5-122B-A10B — JANG_2M vs MLX mixed_2_6

The same-neighborhood comparison retained because JANG is slightly smaller and substantially better.

| Method | GPU memory | MMLU |
| --- | ---: | ---: |
| JANG_2M | 44.7 GB | 79% |
| MLX mixed_2_6 | 45.0 GB | 46% |

Result: JANG is 0.3 GB smaller and wins by +33 MMLU points.

### Mistral-7B — Lower-Bit Coherency Wins

| Prompt family | JANG result | MLX comparison | Outcome |
| --- | --- | --- | --- |
| Photosynthesis | JANG_3M at 3.4 bits | 3.5-bit MLX | JANG gives a coherent explanation; MLX degenerates into number sequences. |
| Arithmetic | JANG_4S at 4.1 bits | 4.5-bit MLX | JANG answers directly; MLX loops the prompt. |

Result: JANG uses fewer bits and keeps coherent output.

### Qwen2.5-3B — Lower-Bit Coherency And MSE Proof

| Comparison | JANG result | MLX comparison | Outcome |
| --- | --- | --- | --- |
| Translation / factual QA | JANG_4S at about 4.1 bits | 4.5-bit MLX | JANG answers directly; MLX echoes or repeats the prompt. |
| Logit MSE | JANG at 3.37 bits | 4.00-bit MLX | JANG has lower measured error while using fewer bits. |

Result: JANG is smaller and stronger on both prompt behavior and measured error.

### SmolLM2-1.7B — Lower-Bit Direct Answer

| Method | Avg bits | Prompt outcome |
| --- | ---: | --- |
| JANG_3M | 3.4 | Answers the spider-legs question directly. |
| MLX 3-bit | 3.5 | Produces number-sequence output. |

Result: JANG uses fewer bits and answers correctly.

### TinyLlama-1.1B — Lower-Bit Topic Retention

| Method | Avg bits | Prompt outcome |
| --- | ---: | --- |
| JANG_4S | 4.1 | Stays on the water-formula question. |
| MLX 4-bit | 4.5 | Derails to a different chemistry question. |

Result: JANG uses fewer bits and avoids topic derailment.

## Removed From Public Model Listing

The page no longer uses a broad “recent uploads” model feed as its primary proof list.
Models are only shown when they pass the stricter smaller-and-decisive-win filter above.

## Runtime Notes

JANG models are designed for native MLX execution on Apple Silicon with compressed weights staying quantized in GPU memory and dequantizing on the fly during inference.

Example:

```bash
pip install jang
jang convert --model Qwen/Qwen2.5-3B --profile JANG_4S --output ./Qwen2.5-3B-JANG_4S/
jang run --model Qwen2.5-3B-JANG_4S.jang
```

## Links

- Homepage: https://jangq.ai/
- Curated model page: https://jangq.ai/models/
- GitHub: https://github.com/jjang-ai/jangq
- Hugging Face account: https://huggingface.co/JANGQ-AI
- vMLX Engine: https://vmlx.net/
- MLX Studio: https://mlx.studio/