Qwen3.5-9B-TNG-PKD-Qwopus-Coder-Fable-Polaris-mxfp4-mlx

Model components

Qwen3.5-9B-TNG-PKD-Qwopus-Coder-Fable

brainwaves
Copy code
arc   arc/e boolq hswag obkqa piqa  wino
bf16      0.643,0.822,0.898,0.718,0.468,0.783,0.707
mxfp8     0.637,0.819,0.892

Qwen3.5-9B-Polaris-PolarisQwen-3NMDST2

brainwaves
Copy code
arc   arc/e boolq hswag obkqa piqa  wino
bf16      0.638,0.829,0.875,0.677,0.450,0.763,0.664
mxfp8     0.628,0.828,0.874,0.668,0.436,0.755,0.672
qx86-hi   0.638,0.832,0.876,0.677,0.446,0.762,0.658

Baseline model

Qwen3.5-9B-Instruct

brainwaves
Copy code
arc   arc/e boolq hswag obkqa piqa  wino
mxfp8     0.571,0.719,0.895,0.683,0.426,0.770,0.671

---

This last merge brings a set of TNG-infused coding sessions recorded in the DS9 Holodeck.

The lessons include:

• Golang coding with Worf and Dax, Nog
• Python coding with Odo and Julian
• Haskell coding with Spock and Data

In all sessions Sisko, Kira, Garak, Quark, Q, and some other DS9 characters provided expert commentary, both in the think tag and in the response.

Along with the lessons, some of the Polaris Alpha questions in the original set were distilled from nightmedia/Qwen3.6-35B-A3B-Holo3-Qwopus-qx86-hi-mlx

The curriculum is exclusively in backend engineering, LLM design, and Holodeck operations(including Quark's Bar and Latinum economy), and was designed with the help of Google Gemini.

-G

---

Model lineage

Qwen3.5-9B-TNG-PKD-Qwopus-Coder-Fable

• nightmedia/Qwen3.5-9B-TNG-PKD-Qwopus-Coder
• armand0e/Qwen3.5-9B-Fable-5-v1

Bringing outstanding Fable traces with tool use and agentic in a very well trained model. Since Fable is XML-friendly, this model reasons better with an XML template--best is the one from DavidAU. Tools work better with json, the current template.

Qwen3.5-9B-TNG-PKD-Qwopus-Coder

• nightmedia/Qwen3.5-9B-Claude-GBO-Fire-Deckard-Agent-Qwopus3.5-Coder-Heretic
• DavidAU/Qwen3.5-9B-Haskell-Rust-Python

This is a set of TNG-infused coding traces, exclusively profiled for Holodeck use, AI training, LLM design, in-memory ops.

Qwen3.5-9B-Claude-GBO-Fire-Deckard-Agent-Qwopus3.5-Coder-Heretic

• nightmedia/Qwen3.5-9B-Claude-GBO-Fire-Deckard-Agent-Heretic
• nightmedia/Qwen3.5-9B-Claude-GBO-Fire-Deckard-Qwopus3.5-Coder-Heretic

Coalesces the previous two models, giving the voice to Agent

Qwen3.5-9B-Claude-GBO-Fire-Deckard-Qwopus3.5-Coder-Heretic

• nightmedia/Qwen3.5-9B-Claude-GBO-Fire-Deckard-Heretic-Thinking
• Jackrong/Qwopus3.5-9B-Coder

Jack managed to build a good model, and since it's different than how the other traces are trained it will raise IQ

Qwen3.5-9B-Claude-GBO-Fire-Deckard-Agent-Heretic

• nightmedia/Qwen3.5-9B-Claude-GBO-Fire-Deckard-Heretic-Thinking
• armand0e/Qwen3.5-9B-Agent

Agent is a collection of quality traces from all over the world, curated by armand0e

• armand0e/badlogicgames-pi-mono-opus-filtered - Pi traces from Claude Opus (mainly 4.5)
• armand0e/kimi-k2.6-claude-code-traces - Claude Code traces from kimi k2.6
• armand0e/kimi-k2.6-agent - Codex traces from kimi k2.6
• armand0e/minimax-m2.7-agent - Pi traces from minimax m2.7
• TeichAI/Claude-Opus-4.6-Reasoning-887x (Downsampled to 200 examples, only present to stabilize chat behavior)

Qwen3.5-9B-Claude-GBO-Fire-Deckard-Heretic-Thinking

• nightmedia/Qwen3.5-9B-Claude-GBO-Fire-Heretic-Thinking
• DavidAU/Qwen3.5-9B-Deckard-Uncensored-Heretic-Thinking

Philip K Dick joins the merge. From here on, there be dragons.

A PKD base will allow multiple realities to develop in a merge.

PKD provides the neural wiring that allows completely different new models to keep joining the NuSLERP.

Since the model is on an ablit base, PKD is at full strength.

Qwen3.5-9B-Claude-GBO-Fire-Heretic-Thinking

• DavidAU/Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-X8b
• DavidAU/Qwen3.5-9B-GBO-Fire-HERETIC-UNCENSORED-THINKING-X8

Excellent Claude trained model joined with a special mix of GLM/Polaris Alpha

Thinking toggle

This model is using an earlier version of the fixed jinja template from froggeric/Qwen-Fixed-Chat-Templates

Drop <|think_on|> or <|think_off|> anywhere in your system or user prompt. The template intercepts the tag, removes it from context so the model never sees it, and flips the mode.

The tag syntax (<|think_on|>, <|think_off|>) uses Qwen's control-token delimiters, so it will never collide with real text. Earlier community templates used /think, which broke legitimate paths like cd /mnt/project/think.

I added a similar set of tags for handling the preserve_thinking flag:

Drop <|think_forget|> or <|think_remember|> anywhere in your system or user prompt to flip the flag.

Model recipe

recipe
Copy code
models:
  - model: DavidAU/Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-X8b
    parameters:
      weight: 1.4
  - model: DavidAU/Qwen3.5-9B-GBO-Fire-HERETIC-UNCENSORED-THINKING-X8
    parameters:
      weight: 0.6
merge_method: nuslerp
dtype: bfloat16
name: Qwen3.5-9B-Claude-GBO-Fire-Heretic-Thinking


models:
  - model: Qwen3.5-9B-Claude-GBO-Fire-Heretic-Thinking
    parameters:
      weight: 1.3
  - model: Qwen3.5-9B-Deckard-Uncensored-Heretic-Thinking
    parameters:
      weight: 0.7
merge_method: nuslerp
dtype: bfloat16
name: Qwen3.5-9B-Claude-GBO-Fire-Deckard-Heretic-Thinking


models:
  - model: Qwen3.5-9B-Claude-GBO-Fire-Deckard-Heretic-Thinking
    parameters:
      weight: 1.6
  - model: armand0e/Qwen3.5-9B-Agent
    parameters:
      weight: 0.4
merge_method: nuslerp
dtype: bfloat16
name: Qwen3.5-9B-Claude-GBO-Fire-Deckard-Agent-Heretic


models:
  - model: Qwen3.5-9B-Claude-GBO-Fire-Deckard-Heretic-Thinking
    parameters:
      weight: 1.6
  - model: Jackrong/Qwopus3.5-9B-Coder
    parameters:
      weight: 0.4
merge_method: nuslerp
dtype: bfloat16
name: Qwen3.5-9B-Claude-GBO-Fire-Deckard-Qwopus3.5-Coder-Heretic


models:
  - model: Qwen3.5-9B-Claude-GBO-Fire-Deckard-Agent-Heretic
    parameters:
      weight: 1.6
  - model: Qwen3.5-9B-Claude-GBO-Fire-Deckard-Qwopus3.5-Coder-Heretic
    parameters:
      weight: 0.4
merge_method: nuslerp
dtype: bfloat16
name: Qwen3.5-9B-Claude-GBO-Fire-Deckard-Agent-Qwopus3.5-Coder-Heretic

models:
  - model: Qwen3.5-9B-Claude-GBO-Fire-Deckard-Agent-Qwopus3.5-Coder-Heretic
    parameters:
      weight: 1.6
  - model: DavidAU/Qwen3.5-9B-Haskell-Rust-Python
    parameters:
      weight: 0.4
merge_method: nuslerp
dtype: bfloat16
name: Qwen3.5-9B-TNG-PKD-Qwopus-Coder

models:
  - model: Qwen3.5-9B-TNG-PKD-Qwopus-Coder
    parameters:
      weight: 1.6
  - model: armand0e/Qwen3.5-9B-Fable-5-v1
    parameters:
      weight: 0.4
merge_method: nuslerp
dtype: bfloat16
name: Qwen3.5-9B-TNG-PKD-Qwopus-Coder-Fable


models:
  - model: Qwen3.5-9B-TNG-PKD-Qwopus-Coder-Fable
    parameters:
      weight: 1.6
  - model: DavidAU/Qwen3.5-9B-Polaris-PolarisQwen-3NMDST2
    parameters:
      weight: 0.4
merge_method: nuslerp
dtype: bfloat16

Use with mlx

bash
Copy code
pip install mlx-lm

python
Copy code
from mlx_lm import load, generate

model, tokenizer = load("Qwen3.5-9B-TNG-PKD-Qwopus-Coder-Fable-Polaris-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, return_dict=False,
    )

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