DavidAU/Qwen3.8-27B-TWIN-TURBO GGUF Released
At a Glance
| Item | Value |
|---|---|
| Repository | DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NM-DAU-NEO-MTP-GGUF |
| Published | 2026-09-10 |
| License | apache-2.0 |
| Formats | GGUF |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code at collection time. Dates are JST.
Overview
DavidAU has released a GGUF version of the custom model based on Qwen3.8-27B, titled “DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NM-DAU-NEO-MTP-GGUF". This model is constructed using multi-stage training and merging techniques with the aim of significantly reducing the number of thinking tokens to anywhere from half to 1/20th of the usual amount while maintaining output detail and quality. It features 5 types of reasoning modes and 5 types of instruct modes that can be switched instantly on APIs or chats, and is built using Unsloth with consumer-grade hardware in mind.
Specifications
- Parameters: 27B class (27.8B)
- Architecture: Qwen3_5ForConditionalGeneration (base model)
- Context Length: 256,144 tokens
Performance
Performance can be checked from the comparison table of various benchmark scores for the base model (Qwen3.8-27B) and others listed on the model card (note that in the table below, comparison target columns are narrowed down to 5 major open-weight models and others).
Text Performance
→ Scroll horizontally to see all columns
| Qwen3.8-27B | Qwen3.6-27B | Qwen3.7-Plus | Muse Glimmer-30B | Opus4.6 Max | |
|---|---|---|---|---|---|
| Coding | |||||
| Agentic terminal coding Terminal Bench 2.1 (Terminus) | 73.0 | 63.4 | 64.0 | 51.7 | 78.2 |
| Agentic coding SWE-bench Pro | 61.7 | 53.5 | 57.6 | 51.2 | 53.4 |
| Repo-level code generation NL2Repo-Bench | 42.3 | 36.2 | 41.1 | — | 47.6 |
| Agentic coding DeepSWE 1.1 | 42.2 | 13.3 | 14.2 | — | — |
| Software engineering QwenSWEBench | 79.0 | 49.3 | 59.2 | — | 63.8 |
| Agent | |||||
| Long-horizon office work CoWorkBench | 70.7 | 61.0 | 65.1 | — | 68.2 |
| Professional job tasks JobBench | 33.4 | 21.8 | 27.6 | — | — |
| Frontier agentic tasks Agents’ Last Exam | Pass@1 20.4 Score 42.9 | Pass@1 10.6 Score 27.3 | Pass@1 13.2 Score 33.6 | — | — |
| General | |||||
| Instruction following IFBench | 79.5 | 69.1 | 79.1 | 77.0 | 62.5 |
| Scientific reasoning GPQA Diamond | 89.2 | 87.8 | 90.3 | 83.5 | 91.3 |
| Multidisciplinary reasoning HLE | 30.8 | 24.0 | 34.7 | 22.0 | 40.0 |
| Competitive coding LiveCodeBench v6 | 90.3 | 83.9 | 89.6 | — | 88.8 |
As performance based on measurements of the base model, it records 73.0 on the coding-related Terminal-Bench 2.1, and shows 61.7 on SWE-bench Pro which fixes real repositories, standing at a high level compared to the same scale band and other models. It marks 89.2 on GPQA Diamond, which measures difficult scientific questions, and 90.3 on LiveCodeBench v6 for competitive programming. On the other hand, it is 30.8 on HLE, which handles extremely difficult questions, showing the characteristics of scores in high-difficulty tasks. In addition, this fine-tuned version is designed to maintain the high-level problem-solving ability and instruction following of the base while suppressing excessive consumption of thinking tokens.
Strengths and Use Cases
This model is built using the multi-stage training and merging techniques “COLD FUSION" and “FABLE FUSION", and is specialized in significantly suppressing excessive thinking (over-reasoning) and token consumption while maintaining the high general intelligence and problem-solving capabilities of the base model. Tags such as creative writing, story, roleplaying, coder, thinking, and reasoning are listed, supporting a wide range of uses including storytelling, creative writing, and coding.
In addition, it is equipped with 5 types of reasoning modes (spoon, einstein, xhigh, medium, low) and 5 types of instruct modes that can be dynamically switched on chats or APIs, making it possible to control behavior according to the situation. Along with light uncensored adjustments, vision capabilities and a maximum context length of 256k tokens are also supported.
How It Differs from Similar Models
While previously introduced models on our site such as the ternary weight model test version prism-ml/Ternary-Bonsai-2-27B-gguf-dev and the Qwen3.8-27B-based ternary weight model Bonsai 2 27B MLX 2bit version released aim for extreme low-bit quantization, lightweighting, and environmental adaptation via ternary weights, this model differs significantly in that it focuses on reducing thinking tokens, a multi-mode switching mechanism, and uncensored adjustments while being based on a consumer-oriented build using Unsloth.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 27.8B parameters (taken from the base model DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 48GB (RTX 6000 Ada / A6000, etc.) | Q5_K_M | 37.9GB | 45.5GB |
| 80GB class (A100 / H100) | IQ3_M | 50.3GB | 60.4GB |
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
How to Get It
It is distributed in GGUF format, and besides standard GGUF quantized versions, various quantization variations are available, such as an “MTP" version supporting Multi-Token Prediction (MTP) and a “TOOLS" version specialized for tool calling. You can download it from the Hugging Face repository using tools like huggingface-cli and run it on standard AI apps and inference engines such as llama.cpp, Ollama, vLLM, and LM Studio. Note that using vision features requires downloading a separate dedicated mmproj file.
Related Articles
- Prism ML Releases Ternary-Bonsai-2-27B-gguf-dev Build
- Signal-3.8-27B-GGUF: Faster and More Token-Efficient
- Prism ML Releases Ternary-Bonsai-2-27B-gguf: A 1.72 bpw Ternary Model
- Salience-27B-R6 GGUF Released by bartowski
Sources
- DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored-NM-DAU-NEO-MTP-GGUF
- DavidAU/Qwen3.8-27B-TWIN-TURBO-Fable-Cold-Fusion-709-L-Uncensored
Update History
- 2026-09-14: Verified the content against the official primary source.
- 2026-09-20: Rewrote the article from re-collected sources and restored it from draft to published.