Sakura DeepSeek-V4.1-Flash-1017 · UD-IQ1_S · Engram Q3_K

The original UD-IQ1_S backbone is unchanged. Only the separate Engram lookup-table files are replaced with Q3_K. All experts and all lookup rows are retained.

No full inference test has been performed. The checks cover integrity, format, addressing and isolated decoder paths. No generation-quality, perplexity, speed or hardware-fit score is claimed for this exact package.

Original versus Sakura

Component Bytes Decimal GB
Matching upstream main GGUF set 232,979,608,475 232.980
Unchanged upstream backbone 122,384,317,883 122.384
Upstream Engram files 110,595,290,592 110.595
Replacement Q3_K Engram files 84,482,513,692 84.483
Sakura main GGUF set 206,866,831,575 206.867
Saved versus this upstream variant 26,112,776,900 26.113 (11.21%)

Main files: 192.660 GiB. Optional DSpark/drafter, vision, imatrix, runtime and documentation are outside this main-model total. File size is not RAM/VRAM consumption or proof that the package fits a particular machine.

What changed and what stayed identical

Part Treatment
4 backbone shards, including tokenizer/model metadata Copied server-side; source sizes and full HF LFS SHA-256 identities retained
Routed experts No pruning, retraining or requantization by Sakura
Engram lookup payload Existing Sakura Q3_K data reused without another quantization pass
Table layout one combined table, layer 1 followed by layer 14; correct split index/count and source tensor-count contract
Actual format GGML type 11 / Q3_K, 110 bytes per 256-value row

The UD-IQ1_S label describes the upstream backbone recipe. Q3_K describes the Engram tables, not the entire model.

Provenance

  • Original checkpoint: deepseek-ai/DeepSeek-V4.1-Flash-1017, revision 1bae0e1088993b9b82de08f59de5be5a34e42170.
  • Direct backbone: unsloth/DeepSeek-V4.1-Flash-1017-GGUF, revision 60ad3651641b5c97679c807e064ab1d7a3ca1455.
  • Q3 payload: Sakura Maya-S Engram Q3_K, frozen input revision 4a701f59f623b08a859368926cd9a210ab446941. It was quantized from peasantsmith's Maya-S Q6_K tables at b9e25f34b53806b24e5b18085de6ed5967ad9ed5, using GGML c7fc5094179bf8a9c73c18593db714f8da41a402.
  • This repack does not requantize the upstream IQ4_NL, Q5_K, Q6_K or MXFP8 table. It reuses the existing Q3 payload and adjusts only the GGUF table header where required.

The official safetensors checkpoint is about 510.297 GB as a broader package; original Maya-S is 339.478 GB. These are context, not scope-matched quality or memory baselines. The matching-upstream row above is this release's reduction baseline.

Table geometry and addressing

Property Value
Engram layers 1 and 14
Layer row counts 384,006,168 / 384,016,682
Total rows 768,022,850
Values per row 256
Total Q3 payload 84,482,513,500 bytes
Storage one combined table, layer 1 followed by layer 14
Combined second-layer starting row, where applicable 384,006,168

Tokenizer-derived token mapping, hash constants, bucket sizes and addressing metadata were compared against the relevant source layout. A 192-row Q6 sample matched Maya and Unsloth before the original table reuse. Maya-XS additionally matched 96 Q6 sample rows and the complete token-map hash. Samples alone do not prove complete original Q6 payload identity.

Download and load

hf download webmp3/Sakura-DeepSeek-V4.1-Flash-1017-Unsloth-UD-IQ1_S-Engram-Q3_K-GGUF --local-dir Sakura-UD-IQ1_S-Engram-Q3_K

Keep all 5 numbered GGUF files together. Load UD-IQ1_S/DeepSeek-V4.1-Flash-1017-UD-IQ1_S-00001-of-00005.gguf with the matching runtime.

Use a runtime supporting this Unsloth DeepSeek-4.1 layout: the inspected danielhanchen/llama.cpp source at b72050d03ae14bb3cb5138ecefab3f3f65157d31 (PR #30817). Its native host-gather Q3_K path was checked on 192 real rows, both layer boundaries, reverse gathers and 64-bit offsets. See the Unsloth guide. This does not qualify arbitrary stock llama.cpp, LM Studio or Maya versions.

What was verified

Check Result and scope
Backbone identity Every copied file's size and LFS SHA-256 matched the pinned source
Table stream Header-only repacks verify complete existing Q3 input hashes and independent Python/native output hashes; see the build record
Published table layout Exact headers, shapes, real types, offsets and split metadata checked
Published real-row samples 192 Q3 rows across both layers, including middle/final rows, compared with frozen fixtures
Decoder and addressing Isolated CPU checks; see compatibility record
Full inference, generation quality, perplexity and speed Not tested for this exact package
GPU execution / Windows / AMD qualification Not performed

Upstream benchmarks and another Sakura variant's results are not measurements of this package. Row reconstruction error is not model quality. Build evidence and exact file identities: build-manifest.json.

Compare the Sakura DeepSeek releases

Open the complete family overview.

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

These compare disk sizes and source recipes, not measured answer quality.

Exact main files

File Bytes SHA-256
UD-IQ1_S/DeepSeek-V4.1-Flash-1017-UD-IQ1_S-00001-of-00005.gguf 5,780,923 b843148a82de2b24df0c5a656669edbb9212eb69a41953249fe6c50d04186177
UD-IQ1_S/DeepSeek-V4.1-Flash-1017-UD-IQ1_S-00002-of-00005.gguf 49,120,285,248 4cb1b13e50db73650d6d9fcaf22cd938b4b2e993740c2abe6815f82107048461
UD-IQ1_S/DeepSeek-V4.1-Flash-1017-UD-IQ1_S-00003-of-00005.gguf 48,899,707,008 def479df0dca997b9cad4716f6375296c1f805887b5cde669105db99f0d9dd75
UD-IQ1_S/DeepSeek-V4.1-Flash-1017-UD-IQ1_S-00004-of-00005.gguf 24,358,544,704 863e1ba47e2479aedc9ce2cc3fe2f57e091548c943bd56d7dd149a771637a17d
UD-IQ1_S/DeepSeek-V4.1-Flash-1017-UD-IQ1_S-00005-of-00005.gguf 84,482,513,692 3c8678a1ae3c25543a8e4ef7f2b157bfbca20cad7f0e52c5c0830a7f0b211037

Credits and license

DeepSeek-AI: original model and Engram. Unsloth / peasantsmith / Project Maya: the credited upstream recipes and runtimes. GGML / llama.cpp: Q3_K format, reference quantizer and decoder. Sakura: table compression, format adaptation, integrity checks and release assembly. No upstream endorsement or authorship of its backbone is claimed.

MIT model/source notices are retained in the included license files. Builds use free public standard CPU runners, no paid runner/GPU or inference API. Public HF storage is best-effort; no quota-neutral deduplication is promised.

中文说明（点击展开）

模型、来源与大小比较

本版本保留 UD-IQ1_S 原始模型主体，所有专家和查找行均保留；仅将独立的 Engram 表文件替换为 Q3_K。Q3_K 只描述查找表，不表示整个模型都是 3 位。

内容 十进制 GB
对应上游主模型分片 232.980
完全不变的模型主体 122.384
上游 Engram 表 110.595
本版本 Q3_K 表 84.483
本版本主分片合计 206.867
减少 26.113

直接主体来源为 unsloth/DeepSeek-V4.1-Flash-1017-GGUF，固定版本 60ad3651641b5c97679c807e064ab1d7a3ca1455。原始模型为 deepseek-ai/DeepSeek-V4.1-Flash-1017，版本 1bae0e1088993b9b82de08f59de5be5a34e42170。Q3 数据来自已有 Sakura Maya-S 表，其输入为 Maya-S Q6_K；本次只适配表头，不再次量化模型主体或 Q3 数据。

完整保留 768,022,850 行，每行 256 值、110 字节。层 1/14 分别有 384,006,168 / 384,016,682 行。主体分片大小及 SHA-256 与上游一致；发布后的表头和 192 行真实 Q3 数据已核对。查找地址、哈希参数及完整 token map 按相应布局检查。

下载时将 5 个编号 GGUF 放在一起，从第一个分片加载。需要支持 Unsloth DeepSeek-4.1 布局的指定 llama.cpp 分支。 固定运行时版本：b72050d03ae14bb3cb5138ecefab3f3f65157d31。详见上方加载说明。

未进行完整模型推理测试。 没有本版本的回答质量、困惑度、速度、GPU 执行、Windows 或 AMD 合格结论。源模型分数不能直接移用；文件大小不等于内存需求。文件校验和、许可、检查边界和其他 Sakura 版本比较见上方英文说明及清单。