merv-qwen3.5-4b-mlx

Merv on Qwen3.5 4B

Qwen3.5 4B (4B parameters), fine-tuned into Merv, a two-headed robot: every answer comes from two personalities at once, each wrapped in its own tag so a chat interface can split the voices.

• Mervin: gloomy and sardonic.
• Mervis: relentlessly cheerful.

User:    What is 2+2?
Mervin:  A trivial sum, naturally assigned to me because apparently no one else in the universe can survive counting to four.
Mervis:  Marvelous! That answer practically sparkles with useful little possibilities, like a sunrise wearing sensible shoes.

Merv is a demonstration of teaching a small model a precise style and output format from a few hundred examples, then running it entirely on the user's own device (the Phi-4-mini version runs in a web browser through WebGPU). The training data (about 260 examples) and the build notebook are at github.com/freeideas/merv. Use follows the base model's licence.

Run it

Weights for Apple Silicon with MLX, quantized to 4 bits.

uvx --from mlx-lm mlx_lm.chat --model freeideas/merv-qwen3.5-4b-mlx

About the author

Made by Carl Free, an AI engineer who trains small, specialized models that beat large general models on narrow tasks at a fraction of the cost. Résumé, demos, and contact: ordinarydata.com/resume. More models: huggingface.co/freeideas.