Sakura — Qwen3.8-27B EfficientThink / GSQ-RCO hybrid (Bonsai-inspired, post-quantization)
Sakura 27B is an experimental post-quantization compressed GGUF based on Qwen3.8-27B-EfficientThink-SimPO-Q2-LynnStyle.gguf.
Provenance
The exact source GGUF is Q2-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q2-LynnStyle.gguf from nerkyor/Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2-GGUF. The foundation model is Qwen/Qwen3.8-27B.
Model information
- Parameters: 26.9B
- Source checkpoint:
Qwen3.8-27B-EfficientThink-SimPO-Q2-LynnStyle.gguf - Source size: 12,999,977,600 Bytes (12.107 GiB)
- Sakura size: 10,672,552,800 Bytes (9.940 GiB)
- Reduction: 17.90%
Measured results
| Metric | Base | Sakura |
|---|---|---|
| WikiText PPL c512 / 8 chunks | 6.6964 | 8.1480 |
| WikiText PPL c4096 / 8 chunks | 6.5008 | 7.5490 |
| ARC-Easy 5-shot | 86.92% | 93.13% |
| MMLU 5-shot | 78.75% | 75.83% |
| Long-context needle | 5/5 | 5/5 |
| Adversarial long-context | 2/2 | 2/2 |
| Prompt processing | 3763 t/s | 3835 t/s |
| Generation | 95.07 t/s | 106.07 t/s |
| VRAM | 13,356 MiB | 11,740 MiB |
| Peak host RAM | 13.1 GB | 10.83 GB |
Speed and memory were measured on an RTX 5090 32 GB using the same setup for Base and Sakura.
Perplexity is worse for Sakura. MMLU is lower by 2.92 points, while ARC-Easy is higher by 6.20 points. These measurements are not a claim of general quality superiority over the base model.
Runtime requirement
Sakura requires the PrismML llama.cpp fork with support for its custom GGUF/runtime operations. Stock/upstream llama.cpp cannot load this GGUF. The verified build at the time of testing was 1a07bfa5f on the prism branch.
Compatibility with LM Studio and Strix Halo has not been tested and is not claimed.
Known limitations
- Perplexity regression relative to the base checkpoint.
- MMLU is 2.92 points lower in the reported 5-shot result.
- This is an experimental release.
- It requires the fork-specific runtime described above.
License
Apache-2.0. The license text is provided in LICENSE.
中文说明 · 樱花 (Simplified Chinese)
English above. 本节为上文的中文翻译(Sakura = 樱花 yīnghuā);完整的独立中文版见 README_zh.md。
Sakura — Qwen3.8-27B EfficientThink / GSQ-RCO 混合版(受 Bonsai 启发,训练后量化)
Sakura 27B 是一个基于 Qwen3.8-27B-EfficientThink-SimPO-Q2-LynnStyle.gguf 的实验性训练后压缩 GGUF。
来源
确切的源 GGUF 是来自 nerkyor/Qwen3.8-27B-EfficientThink-Uncensored-K3-Opus5-Grok4.6-GPT5.6Sol-SFT-SimPO-DFlash2-GGUF 的 Q2-LynnStyle/Qwen3.8-27B-EfficientThink-SimPO-Q2-LynnStyle.gguf。基础模型是 Qwen/Qwen3.8-27B。
模型信息
- 参数量:26.9B
- 源检查点:
Qwen3.8-27B-EfficientThink-SimPO-Q2-LynnStyle.gguf - 源文件大小:12,999,977,600 字节(12.107 GiB)
- Sakura 大小:10,672,552,800 字节(9.940 GiB)
- 缩减:17.90%
测量结果
| 指标 | 基础 | Sakura |
|---|---|---|
| WikiText PPL c512 / 8 chunks | 6.6964 | 8.1480 |
| WikiText PPL c4096 / 8 chunks | 6.5008 | 7.5490 |
| ARC-Easy 5-shot | 86.92% | 93.13% |
| MMLU 5-shot | 78.75% | 75.83% |
| Long-context needle | 5/5 | 5/5 |
| Adversarial long-context | 2/2 | 2/2 |
| Prompt processing | 3763 t/s | 3835 t/s |
| Generation | 95.07 t/s | 106.07 t/s |
| VRAM | 13,356 MiB | 11,740 MiB |
| Peak host RAM | 13.1 GB | 10.83 GB |
速度和内存是在 RTX 5090 32 GB 上测得的,基础版和 Sakura 使用相同的设置。
Sakura 的困惑度更差。MMLU 低 2.92 个百分点,而 ARC-Easy 高 6.20 个百分点。这些测量不是关于整体质量优于基础模型的声明。
运行时要求
Sakura 需要支持其自定义 GGUF/运行时算子的 PrismML llama.cpp 分支。原版/上游 llama.cpp 无法加载此 GGUF。测试时已验证的构建是 prism 分支上的 1a07bfa5f。
与 LM Studio 和 Strix Halo 的兼容性尚未测试,也不作声明。
已知局限
- 相对于基础检查点,困惑度退步。
- 在所报告的 5-shot 结果中,MMLU 低 2.92 个百分点。
- 这是一个实验性发布。
- 它需要上述特定于该分支的运行时。
许可证
Apache-2.0。许可证文本见 LICENSE。