CoolFace
Modelpublic

Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-GGUF

sourceHugging Faceapache-2.0updated 20d agoView on Hugging Face
0likes900downloads
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 NEO-CODER (Official Clean GGUF Suite)</h1>

<h3 align="center">Official Solstice-AI Release &bull; Standard Clean UD 3.0 Matrix &bull; Multi-Token Prediction (MTP) Speculative Tiers &bull; Pure BF16 Multimodal Vision Projector</h3>

<p align="center"> <b>Original Architecture by <a href="https://huggingface.co/Qwen">Qwen / Alibaba Cloud</a> &bull; Uncensored Weights by <a href="https://huggingface.co/DavidAU">DavidAU</a> &bull; Curated &amp; Packaged by <a href="https://huggingface.co/Solstice-AI">Solstice-AI</a></b> </p>

Model Summary

`Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-GGUF` contains the official clean GGUF suite of Qwen3.8-27B NEO-CODER, bringing DavidAU's latest coding and agentic prompt engineering optimizations into standard, clean UD 3.0 GGUF binaries.

Key NEO-CODER Capabilities:

  1. 1.Dynamic Reasoning Effort Controls (`reasoning_effort`):
  2. 2.medium: Suppresses default system prompt injection for direct, unrestricted coding execution and SWE-bench compatibility.
  3. 3.xhigh: Injects deep-reasoning verification tags (<thought>) for complex algorithmic design and proofs.
  4. 4.Deterministic XML Tool Calling: Pre-configured for <tool_call><function=...><parameter=...></function></tool_call> execution.
  5. 5.Pure BF16 Vision Transformer (`mmproj-BF16.gguf`): Standalone 16-bit multimodal vision projector with zero FP16 underflow risks.
  6. 6.Multi-Token Prediction (MTP) Speculative Tiers: Bundles specialized MTP models (speculative-mtp/) and DSpark drafters (speculative/).

File Catalog

FilenamePrecision / QuantSizeRecommended Use Case
`Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-Q4_K_M.gguf`Q4KM (UD-Q4KXL)16.81 GBRecommended: Best balance of speed, RAM footprint &amp; accuracy
`Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-Q5_K_M.gguf`Q5KM19.31 GBHigh-accuracy coding and mathematical reasoning
`Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-Q6_K.gguf`Q6_K21.96 GBNear-lossless weights for complex multi-file refactoring
`Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-Q8_0.gguf`Q8_027.74 GBPure lossless 8-bit precision
*`speculative-mtp/-MTP-Q4KM.gguf`**MTP Q4KM17.23 GBMulti-Token Prediction enabled speculative decoding
*`speculative-mtp/-MTP-Q8_0.gguf`**MTP Q8_028.16 GBLossless MTP speculative decoding
`mmproj-BF16.gguf`Pure BF160.87 GBOfficial standalone Multimodal Vision Projector
`speculative/Qwen3.8-27B-DSpark-Q4_K_M.gguf`DSpark Drafter1.03 GBUltra-fast pre-aligned speculative draft model

Quickstart with llama.cpp

Standard Multimodal Inference:

bash
llama-server \
  -m Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-Q4_K_M.gguf \
  --mmproj mmproj-BF16.gguf \
  -c 131072 \
  --port 8080

Speculative Decoding (1.8x Speedup):

bash
llama-cli \
  -m Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-Q4_K_M.gguf \
  -md speculative/Qwen3.8-27B-DSpark-Q4_K_M.gguf \
  --mmproj mmproj-BF16.gguf \
  -p "Write a high-performance async actor pool in Rust using Tokio."