gemma-4-12B-TNG-V8

This is a Nightmedia release, trained with locally generated content.

This training includes TNG-infused coding lessons.

The teacher model used was the qx86-hi quant of the nightmedia/Qwen3.6-35B-A3B-Holo3-Qwopus-bf16

The lessons are exclusively in backend engineering, LLM design and training, operational Holodeck in Haskell, Rust, Golang and Python.

The set up was in the DS9 Holodeck, with in-person commentary by Spock, Data, Quark, Worf, Odo, Julian, Garak, Q, and many others.

Additionally, some of the Polaris Alpha questions were used to distill from the teacher model.

         arc   arc/e boolq hswag obkqa piqa  wino
bf16     0.553,0.771,0.849
mxfp8    0.551,0.779,0.865
q8-hi    0.552,0.770,0.848
qx86-hi  0.539,0.771,0.835
mxfp4    0.505,0.750,0.853

Quant    Perplexity      Peak Memory   Tokens/sec
bf16     67.645 ± 1.080  30.95 GB      549
mxfp8    74.437 ± 1.177  19.42 GB      373
q8-hi    65.357 ± 1.031  20.54 GB      460
qx86-hi  64.724 ± 1.021  19.00 GB      394
mxfp4   117.640 ± 2.029  13.47 GB      442

Training TNG base

gemma-4-12B-TNG-V6-q8-hi-mlx

         arc   arc/e boolq hswag obkqa piqa  wino
q8-hi    0.524,0.713,0.775

Quant    Perplexity      Peak Memory   Tokens/sec
q8-hi    75.792 ± 1.251  20.54 GB      446

Training 1/4 curriculum

gemma-4-12B-TNG-V1-q8-hi-mlx

         arc   arc/e boolq hswag obkqa piqa  wino
mxfp8    0.439,0.617,0.695
q8-hi    0.475,0.615,0.724
qx86-hi  0.468,0.614,0.753

Quant    Perplexity      Peak Memory   Tokens/sec
mxfp8    85.212 ± 1.319  19.42 GB      360
q8-hi    73.981 ± 1.145  20.54 GB      373
qx86-hi  71.764 ± 1.103  19.00 GB      416

Baseline model

gemma-4-12B-it

         arc   arc/e boolq hswag obkqa piqa  wino
mxfp8    0.385,0.527,0.766,0.509,0.386,0.664,0.579
q8-hi    0.394,0.522,0.774,0.520,0.370,0.674,0.583
qx86-hi  0.391,0.526,0.777,0.521,0.366,0.677,0.590
mxfp4    0.371,0.517,0.660,0.508,0.368,0.677,0.581

Quant    Perplexity      Peak Memory   Tokens/sec
mxfp8   175.766 ± 3.092  19.42 GB      463
mxfp4   283.801 ± 5.361  13.47 GB      484

The V8 iteration introduces a random pick of Claude code traces from angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k-raw mixed in with a random pick of lessons from the first stage.

-G