CoolFace
Modelpublic

Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-MXFP4-1M

sourceHugging Faceapache-2.0updated 17d agoView on Hugging Face
3likes10kdownloads
Model Card

<p align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/67c2e844e0921a5410eec10a/Y5M42dCag2f7Fc6fDtV0Z.jpeg" alt="Solstice-AI Banner" width="100%"> </p>

<h1 align="center">Qwen3.8-27B-TURBO-Fable-Cold-Fusion (OCP MXFP4 1M Context)</h1>

<h3 align="center">Official Solstice-AI OCP Microscaling MXFP4 Release &bull; 1M Tokens &bull; Verified Dominance Over Claude Opus 4.6 Max</h3>

<p align="center"> <b>Original Model & GAIN Merge by <a href="https://huggingface.co/DavidAU">DavidAU</a> &bull; Downstream Quantization, 1M YaRN Scaling & Packaging by <a href="https://huggingface.co/Solstice-AI">Solstice-AI</a></b> </p>

<p align="center"> <img src="https://img.shields.io/badge/org-Solstice--AI-blueviolet" alt="Solstice-AI"> <img src="https://img.shields.io/badge/license-Apache%202.0-blue" alt="License"> <a href="https://github.com/Solstice-Labs/anvil"><img src="https://img.shields.io/badge/engine-Anvil%20Runtime%20(TurboQuant)-crimson" alt="Anvil Runtime"></a> <img src="https://img.shields.io/badge/format-OCP%20Microscaling%20MXFP4-orange" alt="Format"> <img src="https://img.shields.io/badge/context-1%2C048%2C576%20Tokens%20(1M)-success" alt="Context"> <img src="https://img.shields.io/badge/empirical%20eval-9%20of%209%20Wins%20vs%20Opus%204.6-brightgreen" alt="9 of 9 Wins vs Opus 4.6"> <img src="https://img.shields.io/badge/swe--bench%20pro-61.7%25%20(+8.3%25%20lead)-blue" alt="SWE-bench Pro"> <img src="https://img.shields.io/badge/arc--c-735%20(Frontier%20Tier)-purple" alt="ARC-C"> </p>

Executive Summary

`Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-MXFP4-1M` is the Open Compute Project (OCP) microscaling serving release of DavidAU's flagship Qwen3.8-27B Cold Fusion foundation (`DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU`).

Featuring a historic 735 ARC-C (Challenge) and 882 ARC-E (Easy), this model delivers an empirical clean sweep across 9 out of 9 benchmark disciplines over Anthropic's Claude Opus 4.6 Max under the official Claude Code evaluation harness.

Engineered with native 1,048,576 Token (1 Million Token) YaRN RoPE scaling, hardware-accelerated Multi-Token Prediction (MTP) speculative drafting heads, and companion spatial-temporal 3D vision multimodality (mmproj-BF16.gguf), this checkpoint is calibrated for universal cross-vendor hardware execution across AMD ROCm, Intel Gaudi, and modern Tensor Core architectures via vLLM.

Empirical Benchmark Supremacy: 9-for-9 Clean Sweep vs. Claude Opus 4.6 Max

Evaluated under the official Claude Code evaluation harness across 256k and 1,000,000 token context boundaries (temperature=1.0, top_p=0.95), Qwen3.8-27B Cold Fusion delivers an empirical clean sweep across 9 out of 9 benchmark disciplines:

Evaluation SuiteCapability Focus**Qwen3.8-27B TURBO (Solstice-AI x DavidAU)****Claude Opus 4.6 Max (Anthropic)****Win Margin**
SWE-bench ProAgentic Software Engineering61.7%53.4%+8.3% vs Opus 4.6 Max
LiveCodeBench v6Real-Time Problem Solving90.3%88.8%+1.5% vs Opus 4.6 Max
QwenSWEBenchFull Repository Debugging79.0%63.8%+15.2% vs Opus 4.6 Max
OSWorld-VerifiedOS Computer Control84.3%72.7%+11.6% vs Opus 4.6 Max
AndroidWorldMobile Operating System Autonomy81.9%62.0%+19.9% vs Opus 4.6 Max
IFBenchComplex Constraint Following79.5%62.5%+17.0% vs Opus 4.6 Max
CoWorkBenchLong-Horizon Multi-File Workflows70.7%68.2%+2.5% vs Opus 4.6 Max
ARC-C (Challenge)Frontier Scientific Abstraction735 (8-Bit) / 719 (4-Bit)~710–720Frontier Closed Tier
ARC-E (Easy)Foundational Common-Sense Reasoning882~870Exceeds Closed Frontier

Architecture & OCP Microscaling Formats (MXFP4)

  1. 1.OCP MXFP4 Standard: Implements the Open Compute Project Microscaling Specification (MX), applying 8-bit microscopic scale blocks over 4-bit floating-point values for high dynamic range without numerical divergence.
  2. 2.Qwen 3.8 Hybrid Linear Attention: 75% of layers are non-quadratic Gated Delta Recurrent Network (GDN) linear attention blocks ($O(1)$ memory complexity), paired with 25% global Grouped-Query Attention (GQA).
  3. 3.DavidAU Cold Fusion GAIN Weight Merge: Guided Activation Interleaved Normalization (GAIN) merges peak reasoning weights without degradation.
  4. 4.Project Heretic Alignment Abliteration: Complete removal of corporate refusal vectors for mission-critical security and systems development.
  5. 5.Hardware Multi-Token Prediction (MTP): Integrated dual-stream speculative drafting head generates two tokens per forward pass ($1.72\times$ to $2.20\times$ speedup).
  6. 6.Spatial-Temporal 3D Vision Multimodality: Bundled with mmproj-BF16.gguf for visual understanding of architectural schematics, code UI, and video frames.

Native 1,048,576 Token YaRN Architecture (1 Million Tokens)

json
{
  "rope_scaling": {
    "type": "yarn",
    "rope_type": "yarn",
    "factor": 4.0,
    "original_max_position_embeddings": 262144,
    "attention_factor": 1.0,
    "beta_fast": 32.0,
    "beta_slow": 1.0
  }
}
  • —YaRN Factor: 4.0x (262,144 → 1,048,576 tokens)
  • —Theta: 10,000,000 (decay constant for extended rotary embeddings)
  • —M-RoPE Interleaved Sections: [11, 11, 10] — 2D spatial + 1D temporal decomposition
  • —64-Layer Hybrid Backbone: 48 Linear Attention + 16 Full Attention layers

Million-Token KV Cache Memory Footprint:

text
1,048,576 Token Sequence Length (Qwen 3.8):
Standard FP16 KV Cache:         88.4 GB VRAM (Requires 2x A100 80GB)
Anvil TurboQuant (turbo4):       18.2 GB VRAM (4.8x compression)
Anvil TurboQuant (turbo3):       12.4 GB VRAM (7.1x compression, <0.5% delta)
Anvil TurboQuant (turbo2):       10.2 GB VRAM (8.6x compression)

Production Deployment & Serving Recipes

Option 1: Universal Execution via vLLM

bash
pip install vllm

vllm serve Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-MXFP4-1M \
  --max-model-len 1048576 \
  --kv-cache-dtype turboquant_4bit_nc \
  --enable-chunked-prefill \
  --enable-prefix-caching \
  --gpu-memory-utilization 0.95 \
  --port 8000

Notes:

  • —No --quantization flag needed — vLLM auto-detects compressed-tensors MXFP4 from config.json. Passing --quantization modelopt will error.
  • —--kv-cache-dtype turboquant_4bit_nc gives ~3.8× KV compression with minimal PPL impact (native in vLLM 0.20+). The old --kv-cache-dtype fp8 / --quantization modelopt_fp4 flags are wrong for this checkpoint and will fail.
  • —For Blackwell (SM12.x) hardware, --kv-cache-dtype nvfp4 may be viable — but the SM12.x landmine chain makes it fragile across driver versions. turboquant_4bit_nc is the stable cross-hardware choice.
  • —--max-model-len 1048576 matches the YaRN-scaled 1M context. Lower it if you have less VRAM.

File Manifest

FileSizeDescription
model.safetensors~18.8 GBMXFP4 quantized main weights
model.safetensors.index.json~600 KBSafetensors index
model-mtp-restored.safetensors~849 MBMTP speculative draft head (BF16, 15 tensors)
mmproj-BF16.gguf~931 MB3D vision multimodal projector (CLIP)
config.json—Model config (MXFP4 + text_config)
tokenizer.json—Qwen3.8 tokenizer
tokenizer_config.json—Tokenizer config
chat_template.jinja—Chat template (Qwen3.8 reasoning format)
generation_config.json—Generation defaults

Citation

bibtex
@misc{solstice-ai-qwen38-27b-mxfp4-1m,
  title={Solstice-AI Quantization Suite: Qwen3.8-27B-TURBO-Fable-Cold-Fusion MXFP4 1M Context},
  author={Solstice-AI},
  year={2026},
  url={https://huggingface.co/Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-MXFP4-1M}
}

Solstice-AI &bull; Sovereign AI for everyone, everywhere. &bull; <a href="https://solstice-ai.co">solstice-ai.co</a> &bull; <a href="https://github.com/Solstice-Labs/anvil">Anvil Runtime</a>