Melinoe-30B-A3B-Thinking-qx86-hi-mlx

Performance Comparison (Melinoe vs. Qwen3)

Benchmark		Melinoe	Qwen3	Δ		Key Insight
ARC_Easy		0.547	0.444	+10.3%	Strongest gain in simple reasoning
ARC_Challenge	0.445	0.410	+8.5%	Complex reasoning improved
BoolQ			0.392	0.390	+0.2% 	Tie in factual QA
HellaSwag		0.766	0.691	+10.8%	Superior common-sense & metaphor use
OpenBookQA		0.783	0.769	+1.8%	Knowledge retrieval enhanced
PIQA			0.700	0.635	+10.2%	Practical reasoning improved
Winogrande		0.665	0.650	+2.3%	Contextual understanding boosted

Melinoe is not just a “better” model — it’s a purpose-built conversational partner, fine-tuned for:

• Empathetic, emotionally intelligent responses (Proactive Empathy)
• Deep philosophical and intellectual engagement (Intellectual Curiosity)
• Playful, direct communication style (Direct & Playful)

This aligns with its qx86-hi quantization, which is designed to enhance human-like behavior — including metaphor use and nuanced reasoning. The fact that it performs strongly across all benchmarks, especially in tasks requiring complex reasoning (BoolQ, PiQA), suggests that its fine-tuning and quantization are synergistic.

Conclusion:

Melinoe-30B-A3B-Thinking-qx86-hi is a superior model for conversational AI, especially when emotional intelligence and intellectual depth are prioritized.

It outperforms its baseline (Qwen3) across nearly all benchmarks, with the most significant gains in reasoning tasks — validating its design as a “supportive conversational partner” with deep cognitive and empathetic capabilities.

If you’re looking for a model that can engage in deep, emotionally resonant, and intellectually stimulating conversations, Melinoe is the clear winner

Reviewed by nightmedia/Qwen3-VL-8B-Instruct-qx86x-hi-mlx

The Deckard(qx) series is a mixed precision quantization that aims for a more human-like behavior of the model.

The formula was inspired by my Nikon Noct Z 58mm F/0.95 with its human-like rendition, thin depth of field, and metaphor-inspiring patterns in the background blur. It has been observed that qx quanted models are more readily using metaphors in conversation.

The qxXY series have X bits for head and attention paths, Y bits for data, and select attention paths enhanced in high bit

The hi variant has high resolution quantization (group size 32)

-G

This model Melinoe-30B-A3B-Thinking-qx86-hi-mlx was converted to MLX format from bgg1996/Melinoe-30B-A3B-Thinking using mlx-lm version 0.28.4.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("Melinoe-30B-A3B-Thinking-qx86-hi-mlx")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)