Gheorghe Chesler PRO

nightmedia

AI & ML interests

Nightmedia: human-Like AI and the MLX Deckard(qx) Formula Donations are appreciated: BTC:36d7U1n3MFaXgnNRAaEL3Pa3Hy6oFhM7XY

Recent Activity

updated a model 27 minutes ago

nightmedia/Qwen3.6-27B-Miraculix

new activity about 3 hours ago

nbeerbower/Wichtel-Qwen3.6-27B:Miraculix

published a model about 4 hours ago

nightmedia/Qwen3.6-27B-Miraculix

View all activity

Organizations

Posts 8

Post

447

The DavidAU/Qwen3.6-27B-Fable-Fusion-711

This model was assembled at NightmediaAI under the lab name:
nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451
There is ample information on the model card about the model itself.

Here are the base model metrics, as it was merged

Qwen3.6-27B-Architect-Polaris2-Fable-B-F451
          arc   arc/e boolq hswag obkqa piqa  wino
mxfp8     0.711,0.879,0.910,0.790,0.514,0.823,0.763
qx86-hi   0.696,0.876,0.912,0.791,0.518,0.824,0.760
qx64-hi   0.702,0.873,0.909,0.794,0.514,0.822,0.750
mxfp4     0.701,0.873,0.909,0.786,0.488,0.813,0.759

It is a merge of different models, some abliterated, some as they were, so as a result, this source is not Heretic. DavidAU offers the source in Heretic mode here:
DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP

The basic idea behind this model was to build a model that can be used as a base, and has high enough IQ to merge with other models. Every model card shows details of assembly and provides a vibe check with the Nightmedia test prompts.

I merged a few samples here:

with nbeerbower/BigBubba-Qwen3.6-27B
nightmedia/Qwen3.6-27B-Fable-Fusion-711-BigBubba-717-mxfp8-mlx

mxfp8     0.717,0.873,0.912
mxfp4     0.707,0.877,0.912

with migtissera/Tess-4-27B:
nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Tess-mxfp8-mlx

mxfp8     0.712,0.879,0.911,0.792,0.508,0.823,0.764
mxfp4     0.706,0.873,0.910,0.790,0.496,0.817,0.761

with EpistemeAI/Reasoning-Medical-27B (metrics pending):
nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Medical-mxfp8-mlx

mxfp8     0.716,0.877,0.911
mxfp4     0.701

This last model was simply a blind merge, and might not be the proper formula for it, but was entertaining enough to publish as a sample.

I am populating the model card metrics as they become available.

Post

1835

Gemma4 template issues

I ran into this when testing the juiceb0xc0de/locus-gemma-4-e2b, and the response was full of end tags:

> I await your next query, G.<turn|><turn|>><turn|>Your formal request has been processed and analyzed. I am ready to continue the engagement when you are.<turn|>>

It turns out, the fix is in LMStudio:

Why this is critical for Gemma 4 E2B

Gemma 4 models (especially edge variants like E2B-it) utilize Chain-of-Thought thinking layers and structural multi-turn tool schemas natively. The model shifts between channels like <|channel>thought and regular dialogue text seamlessly.

If LM Studio does not explicitly watch for the terminal closing tags as a hard cutoff signal, the edge quantization layers will drop character alignment when processing long context history segments. As a result, the model gets stuck in an infinite state loop, repeating structure summaries instead of returning control back to your prompt session.

cat ~/.lmstudio/config-presets/gemma4.preset.json 
{
  "identifier": "@local:gemma4",
  "name": "gemma4",
  "changed": false,
  "operation": {
    "fields": [
      {
        "key": "llm.prediction.stopStrings",
        "value": [
          "<turn|>",
          "<channel|>",
          "<eos>"
        ]
      }
    ]
  },
  "load": {
    "fields": []
  }
}

That is the exact configuration structure LM Studio requires.

By saving it inside the llm.prediction.stopStrings operational field, LM Studio binds those terminal tokens directly into the underlying runtime client loop rather than the model's architectural blueprint. Every time you load this model profile, the inference wrapper will strictly police and discard those boundary markers before the streaming text token buffer writes to your chat window.

You have successfully stabilized a cutting-edge, programmatic Gemma 4 architecture inside a local GUI environment.

--Gemini

View all Posts

Collections 28

View 28 collections

models 516

datasets 0

None public yet

Gheorghe Chesler PRO

AI & ML interests

Recent Activity

Organizations

Posts 8

Collections 28

nightmedia/granite-4.1-30b-mxfp8-mlx

nightmedia/granite-4.1-30b-mxfp4-mlx

nightmedia/granite-4.1-8b-mxfp8-mlx

nightmedia/granite-4.1-3b-mxfp8-mlx

nightmedia/Qwen3.6-35B-A3B-qx86-hi-mlx

nightmedia/Qwen3.6-35B-A3B-Text-qx64-mlx

nightmedia/Qwen3.6-35B-A3B-Architect-Qwopus-mxfp8-mlx

nightmedia/Qwen3.6-35B-A3B-Holo3-Qwopus-mxfp4-mlx

nightmedia/granite-4.1-30b-mxfp8-mlx

nightmedia/granite-4.1-30b-mxfp4-mlx

nightmedia/granite-4.1-8b-mxfp8-mlx

nightmedia/granite-4.1-3b-mxfp8-mlx

nightmedia/Qwen3.6-35B-A3B-qx86-hi-mlx

nightmedia/Qwen3.6-35B-A3B-Text-qx64-mlx

nightmedia/Qwen3.6-35B-A3B-Architect-Qwopus-mxfp8-mlx

nightmedia/Qwen3.6-35B-A3B-Holo3-Qwopus-mxfp4-mlx

models 516

nightmedia/Qwen3.6-27B-Miraculix

nightmedia/Qwen3.6-27B-Miraculix-mxfp8-mlx

nightmedia/Qwen3.6-27B-GrandWichtel-mxfp8-mlx

nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Medical-mxfp8-mlx

nightmedia/Qwen3.6-27B-Fable-Fusion-711-BigBubba-717-mxfp8-mlx

nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-BigBubba-mxfp4-mlx

nightmedia/Qwen3.5-9B-Holodeck-Fara

nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-Tess

nightmedia/Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-KAT-Coder-C-qx86-hi-mlx

nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451

datasets 0

Gheorghe Chesler PRO

AI & ML interests

Recent Activity

Organizations

Posts 8

Collections 28

models 516 Sort: Recently updated

datasets 0

models 516