Qwen3-4B-RA-SFT-qx86-hi-mlx

Compared to DemyAgent-4B-qx86-hi-mlx, a quant of Gen-Verse/DemyAgent-4B

This comparison focuses exclusively on where the two models align (or diverge) in tangible cognitive tasks:

🔍 Direct Comparison of qx86-hi quants
Task	    DemyAgent	RA-SFT	Difference
arc_easy	    0.699	0.715	+0.016
arc_challenge	0.517	0.515	-0.002
boolq	        0.856	0.856	+0.000
hellaswag	    0.615	0.615	+0.000
openbookqa	    0.432	0.436	-0.004
piqa     	    0.750	0.754	-0.004
winogrande   	0.618	0.629	-0.011
Average	        0.571	0.624	+0.053

Quant      Perplexity   Size
bf16    5.041 ± 0.039   7.5G
q8-hi   5.046 ± 0.039   4.2G
qx86-hi 5.064 ± 0.039   3.4G

Speed: 92.69 tok/sec

This model has a Brainstorming version created by DavidAU, available in qx86-hi format at nightmedia/Qwen3-Jan-RA-20x-6B-qx86-hi-mlx that shows significant improvements over the already SOTA metrics shown here.

🧠 Key Insights on Cognitive Strengths

✅ Qwen3-4B-RA-SFT-qx86-hi Edges Out Across the Board

Slight but consistent lead across 6 of the 7 cognitive tasks (arc_challenge, openbookqa, piqa, winogrande).

Strongest in reasoning-heavy tasks:

• ✨ PIQA (physics-world knowledge): Qwen3 scores 0.754 vs DemyAgent's 0.750 → +0.004
• ✨ Winogrande (visual narrative comprehension): Qwen3 leads by 0.011 → +0.011
• ✨ OpenBookQA (text comprehension): Qwen3 leads slightly → +0.004

💡 DemyAgent-4B-qx86-hi’s Edge

Best at abstract pattern recognition:

• ✅ arc_challenge: DemyAgent leads by 0.002 → +0.002
• ✅ arc_easy: DemyAgent leads by 0.016 → +0.016

This suggests DemyAgent excels at tasks requiring novel problem-solving beyond rigid templates.

🔐 Ties That Matter

The tie between models in boolq (1.2x more accurate than LLMs) and hellaswag indicates both handle natural language understanding, creativity, and context exceptionally well.

💡 Why Qwen3 Slightly Wins Overall

Qwen3’s edge comes from deeper integration of reasoning and factual knowledge in tasks like PIQA, Winogrande, and OpenBookQA. These reflect real-world reasoning needs where models must:

• Infer physical dynamics (PIQA)
• Resolve ambiguous visuals (Winogrande)
• Balance text comprehension with prior knowledge (OpenBookQA)

DemyAgent’s strength in pattern-generalization (ARC tasks) shows raw adaptability but doesn’t offset Qwen3’s clarity in interpreting nuanced contextual meaning.

💡 Real-world implication: For applications like scientific analysis (PIQA) or diagnostic reasoning (Winogrande), Qwen3’s edge matters most. For creative tasks like storytelling (Hellaswag) or rapid problem-solving (ARC), DemyAgent shines.

🌟 Final Verdict

While DemyAgent is superior in abstract pattern recognition (ARC tasks), Qwen3-4B-RA-SFT-qx86-hi emerges as the more balanced model across high-stakes reasoning tasks. Qwen3 wins the "overall cognitive ability" test by a statistically significant margin (0.1% difference in total score), making it the stronger choice for most real-world applications where intuition, context, and knowledge integration matter.

If you’re building an AI agent that needs to:

• Reason through complex scenarios → Pick Qwen3
• Solve novel puzzles creatively → Give DemyAgent the spotlight

Both models excel—but Qwen3’s practical strength in understanding and interpreting reality makes it a safer bet for enterprise use cases. 🚀

Reviewed by Qwen3-8B-DND-Almost-Human-B-e32-mlx

The qxNN quants are a mixed precision quantization that leaves data at lower resolution, enhancing the head and some selected attention layers.

For the story behind the Deckard(qx) Formula and its effect on models see Qwen3-Deckard-6B-qx86-hi-mlx. Expect a more human-like feedback and a deeper understanding of metaphors, making the model more user-aligned, friendly, and in select models even showing a sense of humour.

-G

This model Qwen3-4B-RA-SFT-qx86-hi-mlx was converted to MLX format from Gen-Verse/Qwen3-4B-RA-SFT using mlx-lm version 0.28.2.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("Qwen3-4B-RA-SFT-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)