Gemma-3-27b-it-Gemini-Deep-Reasoning-q8-mlx

The model was trained by DavidAU with TeichAI's gemini-3-pro-preview-high-reasoning dataset.

Comparatively speaking, this model beats all benchmarks seen so far on a small model, judging by test results alone.

It has the best arc numbers in a dense 30B-range model

Gemma-3-27b-it-Gemini-Deep-Reasoning
q8       0.590,0.742,0.883,0.781,0.458,0.822,0.751

This explains the smooth vibe.

A few Nightmedia models for comparison:

Qwen3-30B-A3B-Element7-1M
qx86-hi  0.578,0.750,0.883,0.742,0.478,0.804,0.684

Qwen3-30B-A3B-Element6-1M
qx86-hi  0.568,0.737,0.880,0.760,0.450,0.803,0.714

Qwen3-42B-A3B-Architect
qx86-hi  0.563,0.719,0.881,0.761,0.454,0.805,0.703

Qwen3-32B-Element5-Heretic
qx86-hi  0.483,0.596,0.738,0.754,0.394,0.802,0.710

Qwen3-32B-Engineer4
qx86-hi  0.516,0.661,0.829,0.753,0.386,0.798,0.717

Qwen3-4B-Agent-Claude
qx86-hi  0.572,0.763,0.861,0.708,0.414,0.773,0.676

Qwen3-4B-Engineer3x-F32
qx86-hi  0.613,0.842,0.855,0.748,0.428,0.781,0.709

Qwen3-4B-Engineer3x2
qx86-hi  0.619,0.829,0.850,0.747,0.422,0.776,0.690

Perplexity is usually higher on Gemma compared to Qwen

q8     10.968 ± 0.104
mxfp4  12.381 ± 0.119

This model Gemma-3-27b-it-Gemini-Deep-Reasoning-q8-mlx was converted to MLX format from DavidAU/Gemma-3-27b-it-Gemini-Deep-Reasoning using mlx-lm version 0.30.2.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("Gemma-3-27b-it-Gemini-Deep-Reasoning-q8-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, return_dict=False,
    )

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