Qwen3.8-27B-TURBO NEO-CODER (Official Clean GGUF Suite)

Official Solstice-AI Release • Standard Clean UD 3.0 Matrix • Multi-Token Prediction (MTP) Speculative Tiers • Pure BF16 Multimodal Vision Projector

Original Architecture by Qwen / Alibaba Cloud • Uncensored Weights by DavidAU • Curated & Packaged by Solstice-AI

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:

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

File Catalog

Filename	Precision / Quant	Size	Recommended Use Case
`Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-Q4_K_M.gguf`	Q4_K_M (UD-Q4_K_XL)	16.81 GB	Recommended: Best balance of speed, RAM footprint & accuracy
`Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-Q5_K_M.gguf`	Q5_K_M	19.31 GB	High-accuracy coding and mathematical reasoning
`Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-Q6_K.gguf`	Q6_K	21.96 GB	Near-lossless weights for complex multi-file refactoring
`Qwen3.8-27B-TURBO-Fable-Cold-Fusion-NEO-CODER-Q8_0.gguf`	Q8_0	27.74 GB	Pure lossless 8-bit precision
*`speculative-mtp/-MTP-Q4_K_M.gguf`**	MTP Q4_K_M	17.23 GB	Multi-Token Prediction enabled speculative decoding
*`speculative-mtp/-MTP-Q8_0.gguf`**	MTP Q8_0	28.16 GB	Lossless MTP speculative decoding
`mmproj-BF16.gguf`	Pure BF16	0.87 GB	Official standalone Multimodal Vision Projector
`speculative/Qwen3.8-27B-DSpark-Q4_K_M.gguf`	DSpark Drafter	1.03 GB	Ultra-fast pre-aligned speculative draft model

Quickstart with `llama.cpp`

Standard Multimodal Inference:

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):

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."