Chronos-ENTP-Qwen3.5-9B — MLX 4-bit

A fine-tune of Qwen/Qwen3.5-9B that answers in an ENTP-style voice: direct, curious and a little contrarian. It takes a clear position first, then gives the reasoning. The adapter was merged into the base model and quantized to 4-bit MLX, so it runs on Apple Silicon at about 6 GB.

This is a tone fine-tune only. The training data is about 200 short general-knowledge and opinion Q&A pairs written in the target voice. It adds no new knowledge; factual ability is the base model's.

⚠ Turn thinking OFF

Run this model with enable_thinking=False. With thinking on, it tends not to close its reasoning block and keeps generating until it runs out of token budget, with no answer at the end. With thinking off, it answers in a couple of seconds.

Usage

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("Lifeboat78/Chronos-ENTP-Qwen3.5-9B-MLX-4bit")
prompt = apply_chat_template(processor, model.config,
                             "Is overtime a staffing problem or a routing problem?",
                             num_images=0, enable_thinking=False)
print(generate(model, processor, prompt, max_tokens=300).text)

The model also loads text-only with mlx_lm.

Example

Q: Is it smarter to route techs by geography or by skill?

A: Route by skill first, geography second — a highly skilled tech who's 20 minutes out beats a local tech who can't fix the problem, and the extra travel time is a small price for not sending someone back out.

Training

• Method: LoRA with Unsloth FastVisionModel, fine-tuning the language layers only, so the vision tower is kept as it was. The adapter was then merged and converted to MLX 4-bit (group size 64).
• Data: about 200 synthetic Q&A pairs written to a persona rubric (ENTP traits as described in public MBTI-style profiles), split into train, validation and test sets.

Measured on an M5 MacBook Pro, 24 GB

With thinking off, a 300-token cap and the median of 3 runs: 26.1 tok/s generation, 193 tok/s prompt processing, and a 5.2 GB peak memory.

Limitations

• The voice is opinionated by design, which suits discussion and judgment calls. It is a poor fit for formal writing or anything that needs a neutral tone.
• Image input is kept from the base model but was not tested after fine-tuning.
• It inherits the base model's knowledge limits and failure modes.

License

Apache 2.0, inherited from Qwen3.5-9B.