language:

• en
• zh
  license: apache-2.0
  base_model: DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored
  tags:
• solstice-ai
• davidau
• qwen
• qwen3.8
• qwen3.8-27b
• cold-fusion
• gain
• project-heretic
• heretic
• uncensored
• abliterated
• fable
• cot
• reasoning
• coding
• gguf
• llama.cpp
• ollama
• mtp
• multi-token-prediction
• speculative-decoding
• vision
• multimodal
• mmproj
• q8_0
• q6_k
• q5_k_m
• q4_k_m
• arc-challenge
• 709-arc
  pipeline_tag: image-text-to-text
  datasets:
• Solstice-AI/Solace-1.0-Omni

Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored (GGUF UltraOptimised)

Official Solstice-AI Quantization Suite • Hardware Multi-Token Prediction (MTP) • 10-Level Cognitive Architecture • Twin-Turbo GAIN

Original Model & GAIN Merge by DavidAU • Curated Quantization, MTP Integration & Cognitive Architecture by Solstice-AI

Executive Summary

Solstice-AI/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-GGUF-UltraOptimised-MTP-Curated is the curated, zero-bloat GGUF release of DavidAU's flagship Qwen3.8-27B Twin Turbo Cold Fusion foundation (DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored).

This curated release eliminates non-MAX duplicates, degraded extreme low-bits, and external drafters in favor of pure, Pareto-optimal checkpoints with Native Hardware Multi-Token Prediction (MTP) and Solstice-AI's 10-Level Cognitive Reasoning Architecture.

Curated "GOATed" Checkpoints

Every checkpoint in this suite is a MAX-MTP tier: retaining the critical output tensor (output.weight / lm_head) in unquantized 16-bit precision alongside Q8_0 MTP heads to preserve reasoning accuracy (Arc-C 701+ at 4-bit):

10-Level Cognitive Reasoning Architecture

Built directly into tokenizer_config.json and chat_template.jinja, this suite introduces a 10-level cognitive spectrum. Levels feature soft-elastic pacing (thoughts scale organically to problem difficulty without artificial token caps).

Triggering Modes In-Chat & Via API

• In-Chat Message Tags (works across Ollama, LM Studio, OpenWebUI, LibreChat):
  • Thinking Mode: Add {REASON:<alias>} anywhere in your message (e.g., {REASON:amax}, {REASON:uhigh}, {REASON:athena}). The tag is stripped from the prompt and persists across subsequent chat turns.
  • Instant Instruct Mode (Zero Reasoning Tokens): Prefix with i (e.g., {REASON:iamax}, {REASON:iuhigh}, {REASON:iathena}) to close <think></think> immediately and generate a direct answer framed through that persona.
• API Parameters:

  # Thinking Mode
  response = client.chat.completions.create(
      model="...",
      messages=[{"role": "user", "content": "Analyze system architecture"}],
      extra_body={"chat_template_kwargs": {"reasoning_effort": "amax"}}
  )
  
  # Instant Instruct (0 Thinking Tokens)
  response = client.chat.completions.create(
      model="...",
      messages=[{"role": "user", "content": "Fast code generation"}],
      extra_body={"chat_template_kwargs": {"enable_thinking": False, "reasoning_effort": "uhigh"}}
  )
  

The Cognitive Spectrum

Quickstart

Native MTP Speculative Decoding via llama.cpp

Checkpoints with -MTP- feature native dual-stream token prediction built into the weights (no external drafter file needed):

# High-speed interactive chat with native MTP
llama-cli \
  --hf-repo Solstice-AI/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-GGUF-UltraOptimised-MTP-Curated \
  --hf-file Qwen3.8-27B-TTURBO-Fable-C-Fusion-709-L-Uncen-NM-DAU-NEO-MAX-MTP-Q4_K_M.gguf \
  --mmproj mmproj-BF16.gguf \
  -c 131072 \
  -ngl 99 \
  -p "{REASON:amax} Perform a rigorous architectural evaluation of microservices vs monoliths."

Server Deployment

llama-server \
  --hf-repo Solstice-AI/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-GGUF-UltraOptimised-MTP-Curated \
  --hf-file Qwen3.8-27B-TTURBO-Fable-C-Fusion-709-L-Uncen-NM-DAU-NEO-MAX-MTP-Q4_K_M.gguf \
  --mmproj mmproj-BF16.gguf \
  --host 0.0.0.0 --port 8080 -c 131072 -ngl 99

Citations & Acknowledgments

• DavidAU for the phenomenal Qwen3.8-27B Twin-Turbo Cold Fusion GAIN merged base foundation.
• Qwen Team for the foundational Qwen 3.8 architecture.
• Solstice-AI for downstream curated quantization, MTP packaging, and the 10-level cognitive architecture.