Qwen3-4B-Element8

Brainwaves:

mxfp4    0.533,0.731,0.854,0.689,0.402,0.762,0.657
qx64-hi  0.531,0.728,0.857,0.702,0.410,0.764,0.671
qx86-hi  0.540,0.725,0.866,0.708,0.430,0.769,0.669
bf16     0.542,0.731,0.866,0.706,0.428,0.765,0.655

The last two models merged into Element8:

Qwen3-4B-Element6d

Brings MiniMax-M2.1 traces

mxfp4    0.536,0.718,0.856,0.691,0.400,0.775,0.673
qx64-hi  0.533,0.727,0.865,0.696,0.412,0.766,0.673
qx86-hi  0.536,0.730,0.865,0.704,0.424,0.771,0.665
bf16     0.536,0.731,0.868,0.704,0.434,0.769,0.673

Qwen3-4B-Element7

Brings MiMo-V2-Flash traces

mxfp4    0.532,0.733,0.854,0.690,0.392,0.764,0.661
qx64-hi  0.532,0.729,0.857,0.699,0.414,0.766,0.662
qx86-hi  0.538,0.722,0.864,0.707,0.424,0.768,0.670
bf16     0.540,0.726,0.868,0.707,0.416,0.770,0.657

One of the interesting side effects of the merge is the resistance to quantization: there is almost no difference in cognitive performance betwen quants and full precision, it only shows in openbookqa, exactly where it should, because of the direct dependence on representative depth.

Base models in the merge, showing the highest performing quants in qx86-hi.

The top 4 models are all Nightmedia models, each a successful merge similar to Element8.

Agent               0.603,0.817,0.838,0.743,0.426,0.780,0.708
Agent-Claude        0.561,0.760,0.862,0.714,0.422,0.780,0.683
Engineer-trial17    0.569,0.774,0.849,0.705,0.440,0.773,0.642
Engineer3x-Trial122 0.556,0.757,0.850,0.642,0.436,0.754,0.611

Qwen3-4B-RA-SFT     0.515,0.715,0.856,0.615,0.436,0.754,0.629
Jan-v1-2509         0.435,0.540,0.729,0.588,0.388,0.730,0.633
Polaris-Alpha       0.488,0.653,0.846,0.515,0.378,0.683,0.576
Gemini-Flash        0.386,0.447,0.685,0.582,0.362,0.723,0.593
Apollo-Heretic      0.436,0.583,0.805,0.605,0.398,0.738,0.608
Claude-Haiku-4.5    0.469,0.607,0.842,0.560,0.394,0.705,0.585
Claude-Haiku-4.5-HI 0.416,0.513,0.719,0.581,0.372,0.722,0.603
Gemini-3-Pro        0.466,0.658,0.849,0.568,0.394,0.713,0.571
MiMo-V2-Flash       0.398,0.484,0.673,0.618,0.374,0.721,0.630
MiniMax-M2.1        0.391,0.449,0.627,0.588,0.356,0.718,0.609

These are the Qwen base model performance: where everyone started with their training. The goal of a merge is to stay well above those metrics if it is to have learned something in the process. By merging traces from different sources, the value of truth is dented, but moved towards a safe average, which explains why quants don't mind the squeeze.

Qwen3-4B-Thinking-2507-qx86-hi 0.372,0.414,0.625,0.518,0.366,0.698,0.612
Qwen3-4B-Instruct-2507-qx86-hi 0.447,0.593,0.843,0.448,0.390,0.690,0.554

For coding, reasoning, and high performance agentic, the Engineer, Architect, and Agent series from Nightmedia are hard to beat. Models with less merge steps are usually better at coding, the others better reasoners. Think tags do sometimes appear, when the model hits a hard spot. Those are rare and far apart.

We have models in the same performance envelope as 8B, 14B, 30B, and 6B/8B/13B/21B/42B brainstorming versions by DavidAU

The Element series are experimental, and generally fun as conversational models.

-G

P.S. It seems popular, so I made the source openly available

Stages of development

Qwen3-4B-Engineer

multislerp

janhq/Jan-v1-2509
Gen-Verse/Qwen3-4B-RA-SFT
TeichAI/Qwen3-4B-Instruct-2507-Polaris-Alpha-Distill
TeichAI/Qwen3-4B-Thinking-2507-Gemini-2.5-Flash-Distill

qx86-hi 0.605,0.828,0.843,0.748,0.416,0.777,0.706

Qwen3-4B-Engineer3x

multislerp

Gen-Verse/Qwen3-4B-RA-SFT
TeichAI/Qwen3-4B-Instruct-2507-Polaris-Alpha-Distill
TeichAI/Qwen3-4B-Thinking-2507-Gemini-2.5-Flash-Distill

qx86-hi 0.615,0.835,0.852,0.745,0.420,0.780,0.704

Qwen3-4B-Agent

multislerp, abliterated with heretic by DavidAU:

Qwen3-Engineer3x-4B-Run2-Trial122-7-003
Qwen3-Engineer-4b-run2-trial17-8-004
Qwen3-4B-Apollo-V0.1-4B-Thinking-Heretic-Abliterated

qx86-hi 0.603,0.817,0.838,0.743,0.426,0.780,0.708

Qwen3-4B-Agent-Claude-Gemini-heretic

multislerp, abliterated with heretic by DavidAU:

Qwen3-4B-Agent
TeichAI/Qwen3-4B-Instruct-2507-Claude-Haiku-4.5-Distill
TeichAI/Qwen3-4B-Instruct-2507-Gemini-3-Pro-Preview-Distill
TeichAI/Qwen3-4B-Thinking-2507-Claude-Haiku-4.5-High-Reasoning-Distill

qx86-hi 0.561,0.760,0.862,0.714,0.422,0.780,0.683