---
base_model: Qwen/Qwen3.5-9B
library_name: mlx
pipeline_tag: text-generation
license: apache-2.0
tags:
  - mlx
  - qwen3.5
  - qwenjamin-franklin
  - 4bit
  - local-first
  - text-generation
---

![Qwenjamin Franklin](./assets/qwen9b.png)

# Qwenjamin Franklin 4bit

Qwenjamin Franklin 4bit is the fused Apple Silicon release of the strongest everyday-use branch from the Qwenjamin Franklin workshop line.

It is built from Qwen 3.5 9B and tuned for compact coding help, stricter JSON and tool behavior, and stronger false-premise correction while staying local-first in MLX.

If you want the follow-on release with a more expanded model card, see [`stamsam/Qwenjamin_Franklin_V2`](https://huggingface.co/stamsam/Qwenjamin_Franklin_V2) and its compact sibling [`stamsam/Qwenjamin_Franklin_V2_4bit`](https://huggingface.co/stamsam/Qwenjamin_Franklin_V2_4bit).

## What This Release Is

- Fused MLX 4-bit model
- Base lineage: `Qwen/Qwen3.5-9B`
- Workshop branch lineage: `v14` broad-benchmark daily-driver
- Best fit: Apple Silicon local use where size and speed matter

## Base vs This Model

Internal workshop evals. These scores are project-specific and directional, not public leaderboard claims.

| Eval | Base `Qwen3.5-9B-MLX-4bit` | Qwenjamin Franklin 4bit |
|---|---:|---:|
| `workbench_local_agent_smoke` | 63/100 | 72/100 |
| `full40` | 309/400 | 325/400 |
| `json_hard` | 15/30 | 30/30 |
| `parser_gate` | 2/3, 1/3, 1/3 | 3/3, 3/3, 3/3 |
| `code_smoke` | 95/120 | 95/120 |
| `false_smoke` | 102/110 | 110/110 |
| `tool_schema_canary` | 50/175 | 106/175 |
| `no_tool_leakage` | 99/100 | 100/100 |

## Usage

```bash
python -m mlx_lm generate \
  --model stamsam/Qwenjamin_Franklin_4bit \
  --prompt "Return only valid JSON." \
  --max-tokens 256 \
  --temp 0.0
```

## Notes

- This is the compact Apple Silicon release.
- For strict JSON or code-only tasks, use explicit output instructions in the prompt.
- Verify important outputs before using them in high-stakes workflows.