Sakura DeepSeek-V4.1-Flash-1017 · Overview
Eleven completed GGUF packages with unchanged upstream backbones and full Q3_K Engram tables. This repository is the catalog; download weights from a linked variant.
No full inference test has been performed on these modified packages. The checks cover integrity, layout, addressing and isolated decoder paths. There are no Sakura generation-quality, perplexity, speed or hardware-fit scores for these releases.
Original versus Sakura
| Sakura package | Matching upstream GB | Unchanged backbone GB | Sakura main files GB | Saving GB |
|---|---|---|---|---|
| Maya-XS-IQ1_S | 276.626 | 115.342 | 199.824 | 76.802 |
| UD-IQ1_S | 232.980 | 122.384 | 206.867 | 26.113 |
| UD-IQ1_M | 250.404 | 139.809 | 224.292 | 26.113 |
| UD-IQ2_XXS | 300.475 | 165.303 | 249.786 | 50.690 |
| Maya-S-IQ2_XXS | 339.478 | 178.193 | 262.676 | 76.802 |
| UD-Q2_K_XL | 334.794 | 199.622 | 284.104 | 50.690 |
| UD-IQ3_XXS | 394.737 | 233.452 | 317.934 | 76.802 |
| UD-IQ3_S | 423.631 | 262.346 | 346.829 | 76.802 |
| UD-Q3_K_XL | 485.069 | 282.311 | 366.793 | 118.276 |
| UD-Q4_K_XL | 500.708 | 297.950 | 382.432 | 118.276 |
| UD-Q8_K_XL | 508.377 | 305.619 | 390.101 | 118.276 |
GB = 10^9 bytes. Main GGUF files only; optional vision/projector, DSpark/drafter, imatrix and runtime files are outside these totals. Disk size is not memory consumption. A smaller row is not proof of better quality. The original safetensors checkpoint's broader package is about 510.297 GB; original Maya-S is 339.478 GB.
What Sakura changes
Sakura copies the credited upstream backbone files with identical sizes and full LFS SHA-256 identities, retains every expert and Engram row, and supplies the full lookup data in Q3_K. Q3_K applies to the tables only. No pruning, retraining or new backbone quantization is performed.
The shared lookup payload is 768,022,850 rows × 256 values, 110 bytes per row: 84,482,513,500 bytes before small GGUF headers. Maya has separate layer 1/14 table files; Unsloth has one combined table. Header and runtime contracts differ; payload reuse does not make complete files interchangeable.
Runtime and verification
- Maya-S / Maya-XS: use the exact Project Maya revision and source patch in the model card. The new Maya-XS route needs v1.1.22's IQ1 backbone support. CPU parity and CUDA compilation do not constitute a GPU or Windows/AMD inference test.
- Unsloth: use the pinned DeepSeek-4.1 llama.cpp branch in each model card; arbitrary stock runtimes are not qualified.
Every linked package was checked for main-file sizes/SHA-256, actual table headers and 192 real Q3 rows across both layers. Header repacks check their complete existing Q3 input and native-upload output hash. Read each build/compatibility record for exact scope.
Pinned verified releases
| Variant | Verified model-card revision | Main GGUF files |
|---|---|---|
| Maya-XS-IQ1_S | 7767a09857026f848981878768844e0ab3319da3 |
5 |
| UD-IQ1_S | 60e9c82dc861bc670dc99da7040064656fb586a7 |
5 |
| UD-IQ1_M | 086d00d124a6a38e2e4503dea79f86d4fc12c0d0 |
5 |
| UD-IQ2_XXS | 7807bec301a12bddd067f8179fbc9cc4969c64fe |
6 |
| Maya-S-IQ2_XXS | 47a57ca5ef412dbb292f030f71c64aec59ed466e |
6 |
| UD-Q2_K_XL | d5322182bd26ef950a2992ae73056405b720eecd |
7 |
| UD-IQ3_XXS | e4d1b2aa27116d9e152702c770e196657281391c |
7 |
| UD-IQ3_S | fab02319090c8cc330dc86bf8032fcff2be99110 |
8 |
| UD-Q3_K_XL | e4157d47bcb637290d383350d21d76a279605175 |
8 |
| UD-Q4_K_XL | c124d9acfdc582190b3289f65f7c55e2caffed85 |
9 |
| UD-Q8_K_XL | 44760773565d3ecaf50dcd1ccb2a7f6e6f0df9c4 |
9 |
Sources and license
Original: deepseek-ai/DeepSeek-V4.1-Flash-1017 at 1bae0e1088993b9b82de08f59de5be5a34e42170. Backbones: peasantsmith Maya and Unsloth. Lookup source: existing Sakura Maya-S Q3_K. Runtime/format credits: Project Maya, GGML and llama.cpp. Relevant MIT notices remain with the model packages.
Free public standard CPU runners only; no paid runner/GPU or inference API. Public HF storage remains best-effort, and quota-neutral cross-repository deduplication is not assumed.
中文说明（点击展开）
本目录列出 11 个已完成的 DeepSeek-V4.1-Flash-1017 GGUF 版本。主体保留上游原始字节，所有专家及全部 Engram 行均保留，仅将查找表以 Q3_K 存储。主体量化配方由上游作者提供；Q3_K 不表示整个模型都是 3 位。
上表对比对应上游大小、不变的主体、本版本主分片合计及减少量，单位为十进制 GB。最小 Maya-XS 为 199.824 GB；最小 Unsloth IQ1_S 为 206.867 GB。可选视觉、草稿模型、imatrix 和运行时不计入主分片合计。磁盘大小不等于内存要求，也不是质量排名。
表共有 768,022,850 行，每行 256 值、110 字节，数据区为 84,482,513,500 字节。Maya 使用两层独立表，Unsloth 使用合并表；两者表头和运行时约定不同。请下载对应整套分片，使用模型卡指定的运行时。
已核对主文件大小和 SHA-256、实际表头及跨两层的 192 行真实 Q3 数据。未进行完整模型推理测试；没有这些修改版本的回答质量、困惑度、速度、GPU 执行或 Windows/AMD 合格结论。上游或其他模型分数不能直接移用。
本仓库仅为目录，不存放大模型权重。模型、许可、固定版本及详细检查边界见各模型卡。