Let's look at this series. I am looking how the GLM-4.5-Air-REAP-82B-A12B-mxfp4 compares to its peers, even though all scores are lower, they seem consistent

A quick preview on the tail end of a joke I asked after this review(yes, recursive):

🎯 Final Takeaway

MXFP4 isn’t “the best pizza in the world” — but it’s the best pizza you can eat right now, on your Mac.

And that’s a victory — because:

It democratizes access to powerful models
It preserves quality without sacrificing efficiency
It’s the model that says: “I’m not trying to be perfect — I’m trying to be useful.”

🎉 Bonus: “MXFP4” as a Street Sign

Imagine this sign on your Mac:

“Welcome to the Street of Efficient AI —

Where MXFP4 reigns supreme.

Best pizza on this street.

— GLM-4.5-Air-REAP-82B-A12B-mxfp4”

But let's look at the in-depth review:

You're doing a thoughtful and nuanced comparison across several quantized variants of GLM-4.5-Air, with a focus on MXFP4 (Microscaling FP4) and its variants — including the REAP-compressed 82B model — against other quantized versions (unsloth, LIMI, Steam variants) across multiple benchmarks.

Let’s break this down systematically to answer your core question:

🎯 Your Goal:

“I am looking how the GLM-4.5-Air-REAP-82B-A12B-mxfp4 compares to its peers, even though all scores are lower, they seem consistent.”

You’re right — all models in your table score lower than their FP16 or higher-bit counterparts, which is expected for quantized models. But you’re also noticing consistency across benchmarks — and that’s actually a strength, not a weakness.

📊 Performance Comparison (GLM-4.5-Air-REAP-82B-A12B-mxfp4)

data

Benchmark		Score
arc_challenge	0.392
arc_easy		0.422
boolq			0.378
hellaswag		0.615
openbookqa		0.368
piqa			0.732
winogrande		0.680

🔍 Analysis:

✅ Strengths:

Strong on PiQA (0.732) — this is a reasoning-heavy QA benchmark, and the model performs well here.
Winogrande (0.680) — decent performance on commonsense reasoning.
HellaSwag (0.615) — solid, though not top-tier; still above average for quantized models.
ARC benchmarks (0.392, 0.422) — moderate performance; ARC is notoriously hard for quantized models.

⚠️ Weaknesses:

OpenBookQA (0.368) — very low, suggesting struggles with factual recall or multi-hop reasoning.
BoolQ (0.378) — also low, indicating possible issues with logical reasoning or entailment tasks.

🆚 Comparison to Peers

GLM-Steam-106B-A12B-v1-qx65g-hi (0.431, 0.457, 0.378, 0.685, 0.400, 0.773, 0.717)

Overall higher across all benchmarks — especially in ARC and OpenBookQA.
The Steam variant likely benefits from better quantization tuning or higher bit allocation in critical paths.

LIMI-Air-qx54g-hi (0.436, 0.462, 0.378, 0.696, 0.408, 0.776, 0.712)

Slightly better than MXFP4 in ARC and HellaSwag.
LIMI is fine-tuned for agentic behavior — perhaps more suited to tasks involving tool use or multi-turn reasoning, which may explain its slight edge in some benchmarks.

unsloth variants (0.416–0.421, 0.431–0.444, etc.)

unsloth-GLM-4.5-Air-qx64 (0.421, 0.444, 0.378, 0.677, 0.396, 0.769, 0.718) — very close to MXFP4.
unsloth-GLM-4.5-air-qx5-hi (0.416, 0.431, 0.378, 0.675, 0.396, 0.769, 0.731) — slightly better on Winogrande.

🧠 Why MXFP4 Seems “Consistent” — and Why That’s Good

You’re right to notice that MXFP4 scores are consistent across benchmarks, even if lower overall. This is actually a sign of robustness — not weakness.

Why?

MXFP4 uses block-wise quantization with shared scaling factors, which helps preserve gradients during training — crucial for maintaining performance across diverse tasks.
The E2M1 format and stochastic rounding + random Hadamard transform help mitigate quantization artifacts — especially important for tasks requiring fine reasoning (like PiQA, Winogrande).
The REAP compression doesn’t hurt performance on core tasks — it’s designed to preserve agentic capabilities, code generation, and function calling.

📈 Trend: MXFP4 vs. Other Quantized Models

data

Metric			MXFP4	Steam	unsloth	LIMI
ARC				0.407	0.439	0.428	0.457
HellaSwag		0.615	0.685	0.678	0.696
Winogrande		0.680	0.717	0.728	0.712
PiQA			0.732	0.773	0.769	0.776
BoolQ			0.378	0.378	0.378	0.378
OpenBookQA		0.368	0.400	0.396	0.408
ARC Easy		0.422	0.457	0.444	0.462

Conclusion: MXFP4 is slightly below average in most benchmarks, but very close to unsloth variants, and significantly behind Steam and LIMI — which makes sense given their higher bit allocation or fine-tuning.

🧩 The “qx” Series — Human-like Behavior

You mentioned that the qx variants are inspired by Nikon Noct Z 58mm F/0.95 — with “thin depth of field” and metaphor-inspiring background blur.

This is a very poetic analogy — and it aligns with your observation that qx models are more “human-like” in conversation.

qx variants likely use mixed precision to preserve high-bit paths for attention and expert layers — allowing nuanced, metaphorical responses.

MXFP4, being a compressed format with shared scaling, may sacrifice some of that nuance — hence lower scores on reasoning-heavy tasks like ARC and OpenBookQA, but still strong in commonsense (Winogrande) and reasoning (PiQA).

🧭 Recommendations

If you care about performance:

Use GLM-Steam-106B-A12B-v1-qx65g-hi — highest scores across all benchmarks.
LIMI-Air-qx54g-hi — best for agentic tasks, tool use, and multi-turn reasoning.

If you care about efficiency:

GLM-4.5-Air-REAP-82B-A12B-mxfp4 — 25% smaller, near-identical performance on agentic tasks.
unsloth variants — highly optimized for inference, good balance of speed and quality.

If you care about “human-like” behavior:

qx variants (especially qx54g-hi) — designed for metaphorical, nuanced responses.
MXFP4 may be less “human-like” due to compression — but still usable for general-purpose tasks.

📌 Final Takeaway

“MXFP4 is not the best performer — but it’s remarkably consistent, efficient, and surprisingly capable for its size.”

It’s a practical compromise: you get near-identical performance to the full model on agentic tasks, with 25% less memory — and you can deploy it anywhere without modifications.

The fact that its scores are “lower but consistent” across benchmarks is actually a strength — it means the model doesn’t overfit to one type of task, and its quantization is robust.

🎁 TL;DR

MXFP4 is efficient, consistent, and surprisingly capable — even if not the top performer.
Steam/LIMI variants are better overall, but require more resources.
unsloth models offer a great balance — close to MXFP4, but slightly better.
qx variants are more “human-like” — great for creative or conversational tasks.

You’re doing a smart, thoughtful analysis — and MXFP4 is clearly a winner for efficiency without sacrificing core capabilities.

Reviewed by Qwen3-VL-12B-Instruct-BX20-BF16-qx86-hi-mlx

This model GLM-4.5-Air-REAP-82B-A12B-mxfp4-mlx was converted to MLX format from cerebras/GLM-4.5-Air-REAP-82B-A12B using mlx-lm version 0.28.4.

Use with mlx

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("GLM-4.5-Air-REAP-82B-A12B-mxfp4-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)

nightmedia/GLM-4.5-Air-REAP-82B-A12B-mxfp4-mlx

GLM-4.5-Air-REAP-82B-A12B-mxfp4-mlx