English | 简体中文

Sakura Qwen3.8 Flash-Next Compact · Gyro-M · 512E · CP2

An experimental compact derivative of AgentionAI Gyro-M, retaining all 512 routed experts and their original Gyro encoding. Only the large N-Gram/PLE lookup table is compressed. The matching Windows Vulkan runtime and a relative starter are included.

Size and format

The main model is about 44.5% smaller by file size. There is one GGUF file; no separate MTP draft is required or included. GiB means 2^30 bytes: the main file is approximately 51.01 GB in decimal download units. Runtime and documentation add a small separate download.

TQ2_0 retains the source repository's Gyro-M filename size-class convention. The expert tensors retain AgentionAI's custom trellis encoding; this is not a standard Q3 or standard TQ2_0 whole-model conversion. CP2 identifies the additional table conversion.

Download and start

Keep the GGUF, START.cmd and the complete runtime folder together. Read LICENSE before use. On Windows, double-click START.cmd, wait for the model to load, then open http://127.0.0.1:18098. The same local server provides an OpenAI-compatible endpoint at /v1/chat/completions, model name sakura-gyro-m-cp2. Stop it with Ctrl+C.

The starter resolves paths relative to its own folder, so the entire folder can be moved together. It uses Vulkan, context 8192, one slot, Q8 KV cache, lazy N-Gram reads, and thinking enabled. MTP and the optional CP2 GPU row cache are disabled. Default sampling is temperature 1.0, top_p 0.95, top_k 20, min_p 0.05; the deterministic smoke checks below explicitly used temperature 0/top_p 1 and are not a comparison of sampling presets.

Hardware compatibility

47.5 GiB is the model's file size, not a VRAM requirement. GPU-resident weights, system/shared memory, lazy table I/O, KV cache and buffers have different costs. The tested starter completed on this 64 GB UMA system; no peak-VRAM or universal 48 GB fit claim is made. Increasing context or running other GPU applications changes available memory.

The filename keeps the HF-recognized TQ2_0 label for the hardware widget. A widget grouping does not certify custom-format compatibility or fit; use the included runtime and the measured configuration above. The widget's hardware selection is controlled by each visitor.

Quality and limitations

The completed GGUF, without a pilot overlay, was checked against the unchanged Gyro-M source using the same isolated runtime, corpora and PPL settings. Lower PPL is better.

The original failed Code8 pilot is preserved. The longer, separately planned Code40 expansion used the unchanged encoder and the same <1% criterion; it includes Code8 and is not independent replication. Completed-checkpoint English8 and Code40 reproduce the corresponding overlay PPL to printed precision. Corpus hashes and test scope are recorded in results.json.

In a matched four-case thinking-enabled smoke, both original and CP2 passed all four small cases with nonempty reasoning and normal stops: JSON arithmetic, a Python slice-assignment result, a closed-form triangular function, and integer multiplication. The generated function was checked with a bounded arithmetic AST evaluator, without executing generated code. The relative starter independently answered the Python case correctly with thinking enabled by default.

There are also documented failures: both models returned the same incorrect answers for the Python case with thinking disabled (26 inline and 24 fenced, expected 27). Both returned 27 with thinking enabled. These observations do not isolate model behavior from the shared runtime/template. They are not evidence that all smoke tests passed. Thinking is therefore enabled in the supplied starter.

These are limited integrity/PPL/function checks, not a broad coding benchmark, long-context evaluation, proof of general equivalence, or controlled speed comparison. No MTP, vision, peak-memory or performance claims are imported from the source model card.

Integrity and runtime

All 1,223 non-table tensor payloads were checked against the original source, including expert tensors. Tokenizer and other preserved metadata were checked. Fresh sampled-row re-encoding and CPU/HIP reference checks passed. The complete final GGUF was SHA256-verified after relocation into this folder.

Main GGUF SHA256: 31fa8498ca3084a173ff3d8bad9234f4084a2a525b8e96798b418fa62f5a930d.

See SHA256SUMS for every shipped artifact and runtime/PROVENANCE.json for binary hashes and runtime provenance. The runtime is based on AgentionAI llama.cpp, commit 2353a299b636e7773c98ac496409595a733880f0, with local CP2 support. All comparison and starter tests use byte-identical runtime binaries.

Provenance and license

Base model: Qwen/Qwen3.8-Flash-Next. Direct quantized source: agentionai/Qwen3.8-Flash-Next-Gyro-GGUF, revision 05758344bde211b5b2fd18c1ad5f66f6c2132025, file Qwen3.8-Flash-Next-Gyro-M-TQ2_0.gguf.

The direct quantization parent is AgentionAI Gyro-M. Qwen remains the original upstream base; these ancestry references do not imply endorsement. Credit for Gyro expert encoding, rotations, trellis formats and associated kernels belongs to AgentionAI. Sakura changes only the N-Gram/PLE table representation and its identification metadata, without retraining or pruning experts.

Model weights use Qwen Community License 1.0, not Apache/MIT: read the complete LICENSE, including its commercial-use conditions. Runtime software uses MIT and the third-party notices included in runtime/licenses. Keep NOTICE.md and the applicable license texts when redistributing.

---

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 Flash-Next Compact · Gyro-M · 512E · CP2

这是基于 AgentionAI Gyro-M 的实验性紧凑衍生版本，保留全部 512 个路由专家及其原始 Gyro 编码。仅压缩大型 N-Gram/PLE 查找表。随包提供匹配的 Windows Vulkan 运行时和使用相对路径的启动脚本。

大小与格式

主模型文件体积约减少 44.5%。模型为单个 GGUF 文件；不需要、也未随包提供单独的 MTP 草稿模型。GiB 表示 2^30 bytes，因此主文件按十进制下载单位计算约为 51.01 GB。运行时和文档还会增加少量单独的下载体积。

TQ2_0 沿用源仓库对 Gyro-M 文件名大小档位的标记约定。专家张量保留 AgentionAI 的自定义 trellis 编码；这不是将整个模型转换为标准 Q3 或标准 TQ2_0 格式。CP2 表示额外的查找表转换。

下载与启动

请将 GGUF、START.cmd 和完整的 runtime 文件夹放在一起。使用前请阅读 LICENSE。在 Windows 上双击 START.cmd，等待模型加载，然后打开 http://127.0.0.1:18098. 同一个本地服务在 /v1/chat/completions 提供 OpenAI 兼容接口，模型名称为 sakura-gyro-m-cp2。按 Ctrl+C 停止服务。

启动脚本相对于自身文件夹解析路径，因此整个文件夹可以一起移动。它使用 Vulkan、8192 上下文、单个处理槽位、Q8 KV 缓存、按需读取 N-Gram 表，并且启用思考模式。MTP 和可选的 CP2 GPU 行缓存均关闭。默认采样参数为 temperature 1.0、top_p 0.95、top_k 20、min_p 0.05；下文的确定性功能检查明确使用 temperature 0/top_p 1，并非对不同采样预设的比较。

硬件兼容性

47.5 GiB 是模型文件大小，不是显存需求。 GPU 常驻权重、系统／共享内存、查找表按需 I/O、KV 缓存和缓冲区各有不同开销。测试用启动配置在这台 64 GB UMA 系统上完成了运行；并未声称峰值显存占用，也未声称所有 48 GB 配置都能容纳。增大上下文或运行其他 GPU 应用会改变可用内存。

文件名保留了 HF 硬件组件可识别的 TQ2_0 标签。组件中的分组不能证明自定义格式的兼容性或内存适配情况；请使用随包运行时和上述实测配置。硬件组件的具体硬件选择由每位访问者自行控制。

质量与局限

测试使用完整的最终 GGUF，未使用试点叠加覆盖，并在相同的隔离运行时、语料和 PPL 设置下与未修改的 Gyro-M 源模型比较。PPL 越低越好。

原始 Code8 试点未通过的结果仍予保留。另行计划的较长 Code40 扩展测试使用同一个未修改的编码器和相同的 <1% 标准；它包含 Code8，因此并非独立重复验证。完整检查点上的 English8 和 Code40 在已打印的数值精度内复现了各自叠加覆盖测试的 PPL。语料哈希和测试范围记录在 results.json 中。

在条件匹配的四项启用思考模式的功能检查中，原始模型和 CP2 均通过全部四个小案例，具有非空的推理内容并正常停止：JSON 算术、Python 切片赋值结果、三角数闭式函数和整数乘法。生成的函数通过受限的算术 AST 求值器检查，未执行生成的代码。相对路径启动脚本还在默认启用思考模式的情况下独立正确回答了该 Python 案例。

同时也有已记录的失败案例：关闭思考模式时，两个模型对同一个 Python 案例都给出了相同的错误答案（行内形式为 26，代码块形式为 24，预期为 27）。启用思考模式后，两者均回答 27。这些观察无法将模型自身行为与共享运行时／模板的影响区分开，也不能证明所有功能检查都已通过。因此，随包启动脚本默认启用思考模式。

这些只是有限的完整性、PPL 和功能检查，并非广泛的代码能力基准、长上下文评估、一般等价性证明或受控速度比较。没有将源模型卡中的 MTP、视觉、峰值内存或性能宣称作为本版本自己的结果。

完整性与运行时

已将全部 1,223 个非查找表张量的数据区域与原始源文件逐一比较，其中包括专家张量。已检查分词器及其他应保留的元数据。重新抽样行编码和 CPU/HIP 参考检查通过。将最终 GGUF 移入本文件夹后，已对完整文件验证 SHA256。

主 GGUF 的 SHA256：31fa8498ca3084a173ff3d8bad9234f4084a2a525b8e96798b418fa62f5a930d。

所有随包文件的校验值见 SHA256SUMS；二进制哈希和运行时来源见 runtime/PROVENANCE.json。运行时基于 AgentionAI llama.cpp，提交为 2353a299b636e7773c98ac496409595a733880f0，并加入本地 CP2 支持。全部对比测试和启动脚本测试使用逐字节一致的运行时二进制文件。

来源与许可证

基础模型：Qwen/Qwen3.8-Flash-Next。直接量化来源：agentionai/Qwen3.8-Flash-Next-Gyro-GGUF，修订为 05758344bde211b5b2fd18c1ad5f66f6c2132025，文件为 Qwen3.8-Flash-Next-Gyro-M-TQ2_0.gguf。

直接量化父模型为 AgentionAI Gyro-M。Qwen 仍是原始上游基础模型；这些来源引用不表示上游为本版本背书。Gyro 专家编码、旋转、trellis 格式及相关内核的贡献属于 AgentionAI。Sakura 仅修改 N-Gram/PLE 查找表的表示方式及其标识元数据，没有重新训练，也没有裁剪专家。

模型权重采用 Qwen Community License 1.0，并非 Apache/MIT；请阅读完整的 LICENSE，包括其中的商业使用条件。运行时软件采用 MIT 及 runtime/licenses 中随附的第三方许可声明。再次分发时，请保留 NOTICE.md 和相应许可证文本。