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
Executive Summary
Solstice-AI/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-UltraOptimised-MTP-1M 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):
| Checkpoint File | Size | VRAM Budget | Optimal Target |
|---|---|---|---|
...-MAX-MTP-Q4_K_M.gguf | 17.23 GB | 16 GB VRAM | The GOAT 4-bit (16-bit lm_head, Arc-C 701, ideal for RTX 4080 / 16GB GPUs) |
...-MAX-MTP-Q5_K_M.gguf | 19.73 GB | 24 GB VRAM | The GOAT 5-bit (sweet spot of generation speed & reasoning depth) |
...-MAX-MTP-Q6_K.gguf | 22.38 GB | 24–32 GB VRAM | High-fidelity near-lossless sweet spot for RTX 3090/4090 & Apple Silicon |
...-MAX-MTP-Q8_0.gguf | 28.16 GB | 32 GB+ VRAM | Full reference precision |
mmproj-BF16.gguf | 0.87 GB | System RAM/VRAM | Spatial-temporal multimodal vision projector (images & video frames) |
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.
- Thinking Mode: Add
- 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
| Level | Primary Key | Technical Aliases | Mythological Alias | Cognitive Framework & Behavior |
|---|---|---|---|---|
| 0 | disabled | none, off, direct | Mortal | 0 tokens: <think></think> closed immediately for instant direct output. |
| 1 | ulow | ultra-low, micro | Hermes | Rapid instinct & sanity check. Direct path from premise to verdict (<150 tokens). |
| 2 | low | compact, fast | Apollo | Crisp logic and premise validation with zero cognitive overhead. |
| 3 | lmed | low-medium, targeted | Artemis | Boundary hunter: tests zero conditions, nulls, and hidden edge cases. |
| 4 | medium | med, balanced | Athena | Strategic balance: evaluates architectural trade-offs and structural cohesion. |
| 5 | mhigh | medium-high, architect | Prometheus | Proactive forethought: models 10x/100x scale, failure modes, and fault tolerance. |
| 6 | high | deep, thorough | Solstice | Deep systemic derivation: multi-branch hypothesis trace and red-team falsification. |
| 7 | xhigh | extreme-high (Default) | Hyperion | Native Qwen 3.8 continuous derivation and exhaustive semantic deconstruction. |
| 8 | uhigh | ultra-high, swarm | Einstein | 20-Agent Swarm: Deploys 20 virtual perspective agents across Sternberg styles. |
| 9 | amax | absolute-max, deep-research | Oracle | Deep Research Council: 1–5 complexity scaling, multi-expert panel & audit matrix. |
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-UltraOptimised-MTP-1M \
--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 1048576 \
-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-UltraOptimised-MTP-1M \
--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 1048576 -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.