Add files using upload-large-folder tool

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128-DSpark-support.gguf filter=lfs diff=lfs merge=lfs -text
+DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf filter=lfs diff=lfs merge=lfs -text

BUILD_MANIFEST.json

@@ -0,0 +1,1046 @@
+{
+  "schema": "deepseek-model-tools.build128.v1",
+  "artifact": "ds4-quality128",
+  "output": "DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128",
+  "engine": "ds4",
+  "dspark": true,
+  "created_at": "2026-08-01T13:29:21.384605+00:00",
+  "source": {
+    "path": "apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8",
+    "only_weight_source": true,
+    "config_sha256": "6c8f3d2d3b48707541b88f32f22ef3f0f8a6b57d8523281e2b8d3cdb0ae9a023",
+    "index_sha256": "98efab455cf08dfbbbaaba6f570e1bf10bf927d2b4c3c453a59c2f6f0e3be92b",
+    "abliteration_manifest_sha256": "8e6f40f3d92542720aabdafb728f90e3397fba68c6a8d7ba91772b5dd41bc767",
+    "edited_tensors": 36,
+    "upstream_model": "deepseek-ai/DeepSeek-V4-Flash-0731",
+    "upstream_revision": "9e165c30e2704aec5d9d593cce3eebd58bbef1cb",
+    "weight_shards": [
+      {
+        "filename": "model-00001-of-00048.safetensors",
+        "bytes": 1059061856,
+        "sha256": "f3668ba4cccf1ca6a7eb84e888fb92c1cdc7204d472ba9db771e6fd3abf6b874",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00002-of-00048.safetensors",
+        "bytes": 3566321192,
+        "sha256": "77b26c939a0e25b3113c8d6bb04e1901a748bd4a7d2589e3bfdaabdf1e9bba14",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00003-of-00048.safetensors",
+        "bytes": 3566321192,
+        "sha256": "412abf4c906faadc221ef0cb50f90fe20bde8454a08ad4dc2364b6b79e7fda5c",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00004-of-00048.safetensors",
+        "bytes": 3596229272,
+        "sha256": "9610f56bc587fb0ff9a8b68a60299482ee8c433fe5b5587e4257aca98add4a2e",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00005-of-00048.safetensors",
+        "bytes": 3568768976,
+        "sha256": "f87a5ac7b8becc31f9c3169afd3a6f33fb82b4af9e21022e3755a10bc28f0180",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00006-of-00048.safetensors",
+        "bytes": 3590024776,
+        "sha256": "4a4f3764e3fc772b9fba67f0a44ef68e18f178b6f00faa80b75db549e51894cd",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00007-of-00048.safetensors",
+        "bytes": 3568768976,
+        "sha256": "df81bb80e27a689e01fa579eebd6499f86e0b6105f7fea18961aa5eebbbee9bc",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00008-of-00048.safetensors",
+        "bytes": 3590024776,
+        "sha256": "224968d2b27f8669365ec08657a768dfec40da0585f85f302a31495931f6a526",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00009-of-00048.safetensors",
+        "bytes": 3568768976,
+        "sha256": "04d69ef1071fff8721c62968c200a5583122b59b015e9ef9b2978bfed271b2b7",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00010-of-00048.safetensors",
+        "bytes": 3590024776,
+        "sha256": "627145f4ebeb1cc3f5bdd03416b8cb7370b3c96974853cbeb8e5516ad5713e49",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00011-of-00048.safetensors",
+        "bytes": 3568768976,
+        "sha256": "e4b8e601dcbebe902e0102e7b098b670a121cb8b9564dd719fc41d782c8416e0",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00012-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "d200bcd9186f15344b8271f845bd81f782509c5997ee493b1a96079e91a1a8ac",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00013-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "9c36e454195177bafdb866bbfd052f69dd727fafbe328c0d5d4d659ab49389b2",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00014-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "ce44ce426be0ed8b21141ac1659ca108d2d0705b5b442a7691511bc38831bf9e",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00015-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "f9b40ffad55e01cbfc012c737640f882ad7161026ea1c6d878d272c655e74b02",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00016-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "acf5cbaf18a2d2e005358b0a81527ed70bc92f352ab46f658172a0505983bb2c",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00017-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "d67716007b48845ed1de1473ca0ab15908a9dff8b592bb62ab95137cc67cf198",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00018-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "ac550ef17db3b8c95c2a7858fefa7ba08c94b5e6feafa8bfdb3f9763612fd3b1",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00019-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "c31a222ce5c7e0d02b8c2614badd16781cb3979f7e516d5293ab976ee670f34f",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00020-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "a27655e314a1919d9c7042b1c7f842fa8c8f1bf77df4c0a9586371b4ab3029d9",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00021-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "4068e1c0a59dcd4b7a22c09cf870e86439407f1eb8c8669b20d81d3294e4ca60",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00022-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "eb87aef24cbf17c98f8ab4803590606db89acc69c83b9d4f44919bac328d830d",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00023-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "4b8d427c9d2f285986d18b465f3ffcb73c12a109e1dedc8b0a3b6ac83e0e1ab7",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00024-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "2cf277a128b4f51dcd13c130db18e325006962eec0a1f1e90596d32085980daa",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00025-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "dd813e4837baa77bd5fac84d5dce225e6e949c7e781cf40621c06081b6bfcbe1",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00026-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "a4c8f53dbf43374c5a277e992973b0f5b9e916022dd17cecc024c4290165eae8",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00027-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "70050f03122a0050bad34b03eb0bb88def42f2c3ea99c525c1af776409efb76e",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00028-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "51113c3fce251a429452829adea388497184a80e0c0023febbb88c4068f65842",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00029-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "f7b9198bb1e6c05ab5e88bf22fbe78d3f67f789a494d4c1be763787ee11f5f1e",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00030-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "51e9984913ff3abbb7649d8738e764360c165401a9c31d9b53715c1fbe313061",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00031-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "d31d4f707391eca341eaa6025f3e8ecf21ee254340bd8f11d494200aa8923447",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00032-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "b7ecad0122f051d1e2a0953312640ed53ac7da3875766e1e70ff0d0666c28087",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00033-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "7bf5a6c9a2774698ebf819f3abf20604376fd75daecb6ecb6499df964e4447af",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00034-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "b0de49f523078e8766c5deb61be4fc73da35271f3bd354bb3e7c71e975cd8be5",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00035-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "471a793bb28aa16c95d0cad6e9194191fd9c2e0657c9e35917bd7856ddbf8187",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00036-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "002bdca78f533708438ed04ca42139fb67f0325571adb29645f5ef38be54b066",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00037-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "71f148a710c0ff1b505d6e285b1e782eef87d14d5fa4f5411e1faa4c06210708",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00038-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "ab90e440639d4b0de9a9e2f7a48238e7ccdc08c37d40cf91d1a380ba26c1938a",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00039-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "26f3acc9d23d426edabef80260ac4030aad030553356c47b1d5d47667c0813c3",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00040-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "549d59fc74aa7bba55678da85e5a369ebe8017763c4d5f83359e15323659afcf",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00041-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "1868317b4ff748d1388805c530fc7b0310397fdef8ac1c0bd7f034f14abe83ce",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00042-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "11e81dc83b2f6ab688b9205552a147cf781d73ee791a8d54c763e0a461006def",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00043-of-00048.safetensors",
+        "bytes": 3568770544,
+        "sha256": "bc1dc626c7ec234f58807443ecbdec2a5927be771339697cbac3dca94c6ed9a0",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00044-of-00048.safetensors",
+        "bytes": 3590026352,
+        "sha256": "a20dfe1e58e2744ed8c82cdca63326ed62c2f024361a072ecb827edd9e0287bc",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00045-of-00048.safetensors",
+        "bytes": 1059332516,
+        "sha256": "a5be6aed7b84fc87ec42b5d24ba0b0d67f253a3906fcd99c13f4f7be5958fc00",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00046-of-00048.safetensors",
+        "bytes": 3610455184,
+        "sha256": "da1e0108b463653b292a648aef58b4dbbeac03a8c213c53829b99c6d84d941c9",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00047-of-00048.safetensors",
+        "bytes": 3560111960,
+        "sha256": "b8b41236c1ccf4378a3ae2ccc22a784ca1a5681f13a1d64c262836b788a0f093",
+        "upload_mode": "lfs"
+      },
+      {
+        "filename": "model-00048-of-00048.safetensors",
+        "bytes": 3692775244,
+        "sha256": "d8b68325b29a33f7ee43cafcaa69c5de8c1de7c1f6007758961ad4ed11aa74db",
+        "upload_mode": "lfs"
+      }
+    ],
+    "weight_shard_bytes": 166886535336,
+    "weight_shard_hashes_verified": true,
+    "weight_shard_hashes_verified_at": "2026-08-01T11:53:46.973992+00:00",
+    "post_hash_stat_snapshot": {
+      "model-00001-of-00048.safetensors": {
+        "bytes": 1059061856,
+        "mtime_ns": 1785506342100000000,
+        "ctime_ns": 1785506342100000000,
+        "device": 16777242,
+        "inode": 10534358
+      },
+      "model-00002-of-00048.safetensors": {
+        "bytes": 3566321192,
+        "mtime_ns": 1785506344340000000,
+        "ctime_ns": 1785506344340000000,
+        "device": 16777242,
+        "inode": 10538402
+      },
+      "model-00003-of-00048.safetensors": {
+        "bytes": 3566321192,
+        "mtime_ns": 1785506346560000000,
+        "ctime_ns": 1785506346560000000,
+        "device": 16777242,
+        "inode": 10552010
+      },
+      "model-00004-of-00048.safetensors": {
+        "bytes": 3596229272,
+        "mtime_ns": 1785506349630000000,
+        "ctime_ns": 1785506349630000000,
+        "device": 16777242,
+        "inode": 10565618
+      },
+      "model-00005-of-00048.safetensors": {
+        "bytes": 3568768976,
+        "mtime_ns": 1785506351960000000,
+        "ctime_ns": 1785506351960000000,
+        "device": 16777242,
+        "inode": 10579340
+      },
+      "model-00006-of-00048.safetensors": {
+        "bytes": 3590024776,
+        "mtime_ns": 1785506354260000000,
+        "ctime_ns": 1785506354260000000,
+        "device": 16777242,
+        "inode": 10592957
+      },
+      "model-00007-of-00048.safetensors": {
+        "bytes": 3568768976,
+        "mtime_ns": 1785506356530000000,
+        "ctime_ns": 1785506356530000000,
+        "device": 16777242,
+        "inode": 10606655
+      },
+      "model-00008-of-00048.safetensors": {
+        "bytes": 3590024776,
+        "mtime_ns": 1785506358750000000,
+        "ctime_ns": 1785506358750000000,
+        "device": 16777242,
+        "inode": 10620272
+      },
+      "model-00009-of-00048.safetensors": {
+        "bytes": 3568768976,
+        "mtime_ns": 1785506362250000000,
+        "ctime_ns": 1785506362250000000,
+        "device": 16777242,
+        "inode": 10633970
+      },
+      "model-00010-of-00048.safetensors": {
+        "bytes": 3590024776,
+        "mtime_ns": 1785506365890000000,
+        "ctime_ns": 1785506365890000000,
+        "device": 16777242,
+        "inode": 10647587
+      },
+      "model-00011-of-00048.safetensors": {
+        "bytes": 3568768976,
+        "mtime_ns": 1785506369620000000,
+        "ctime_ns": 1785506369620000000,
+        "device": 16777242,
+        "inode": 10661285
+      },
+      "model-00012-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506377140000000,
+        "ctime_ns": 1785506377140000000,
+        "device": 16777242,
+        "inode": 10674902
+      },
+      "model-00013-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506386820000000,
+        "ctime_ns": 1785506386820000000,
+        "device": 16777242,
+        "inode": 10688602
+      },
+      "model-00014-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506396120000000,
+        "ctime_ns": 1785506396120000000,
+        "device": 16777242,
+        "inode": 10702221
+      },
+      "model-00015-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506405630000000,
+        "ctime_ns": 1785506405630000000,
+        "device": 16777242,
+        "inode": 10715921
+      },
+      "model-00016-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506415210000000,
+        "ctime_ns": 1785506415210000000,
+        "device": 16777242,
+        "inode": 10729540
+      },
+      "model-00017-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506424840000000,
+        "ctime_ns": 1785506424840000000,
+        "device": 16777242,
+        "inode": 10743240
+      },
+      "model-00018-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506432280000000,
+        "ctime_ns": 1785506432280000000,
+        "device": 16777242,
+        "inode": 10756859
+      },
+      "model-00019-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506439960000000,
+        "ctime_ns": 1785506439960000000,
+        "device": 16777242,
+        "inode": 10770559
+      },
+      "model-00020-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506449550000000,
+        "ctime_ns": 1785506449550000000,
+        "device": 16777242,
+        "inode": 10784178
+      },
+      "model-00021-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506459270000000,
+        "ctime_ns": 1785506459270000000,
+        "device": 16777242,
+        "inode": 10797878
+      },
+      "model-00022-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506468940000000,
+        "ctime_ns": 1785506468940000000,
+        "device": 16777242,
+        "inode": 10811497
+      },
+      "model-00023-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506478610000000,
+        "ctime_ns": 1785506478610000000,
+        "device": 16777242,
+        "inode": 10825197
+      },
+      "model-00024-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506488130000000,
+        "ctime_ns": 1785506488130000000,
+        "device": 16777242,
+        "inode": 10838816
+      },
+      "model-00025-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506495770000000,
+        "ctime_ns": 1785506495770000000,
+        "device": 16777242,
+        "inode": 10852516
+      },
+      "model-00026-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506505150000000,
+        "ctime_ns": 1785506505150000000,
+        "device": 16777242,
+        "inode": 10866135
+      },
+      "model-00027-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506514650000000,
+        "ctime_ns": 1785506514650000000,
+        "device": 16777242,
+        "inode": 10879835
+      },
+      "model-00028-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506524000000000,
+        "ctime_ns": 1785506524000000000,
+        "device": 16777242,
+        "inode": 10893454
+      },
+      "model-00029-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506533430000000,
+        "ctime_ns": 1785506533430000000,
+        "device": 16777242,
+        "inode": 10907154
+      },
+      "model-00030-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506540820000000,
+        "ctime_ns": 1785506540820000000,
+        "device": 16777242,
+        "inode": 10920773
+      },
+      "model-00031-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506550210000000,
+        "ctime_ns": 1785506550210000000,
+        "device": 16777242,
+        "inode": 10934473
+      },
+      "model-00032-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506559630000000,
+        "ctime_ns": 1785506559630000000,
+        "device": 16777242,
+        "inode": 10948092
+      },
+      "model-00033-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506567220000000,
+        "ctime_ns": 1785506567220000000,
+        "device": 16777242,
+        "inode": 10961792
+      },
+      "model-00034-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506576510000000,
+        "ctime_ns": 1785506576510000000,
+        "device": 16777242,
+        "inode": 10975411
+      },
+      "model-00035-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506585920000000,
+        "ctime_ns": 1785506585920000000,
+        "device": 16777242,
+        "inode": 10989111
+      },
+      "model-00036-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506595350000000,
+        "ctime_ns": 1785506595350000000,
+        "device": 16777242,
+        "inode": 11002730
+      },
+      "model-00037-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506602780000000,
+        "ctime_ns": 1785506602780000000,
+        "device": 16777242,
+        "inode": 11016430
+      },
+      "model-00038-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506611980000000,
+        "ctime_ns": 1785506611980000000,
+        "device": 16777242,
+        "inode": 11030049
+      },
+      "model-00039-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506621720000000,
+        "ctime_ns": 1785506621720000000,
+        "device": 16777242,
+        "inode": 11043749
+      },
+      "model-00040-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506630860000000,
+        "ctime_ns": 1785506630860000000,
+        "device": 16777242,
+        "inode": 11057368
+      },
+      "model-00041-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506638120000000,
+        "ctime_ns": 1785506638120000000,
+        "device": 16777242,
+        "inode": 11071068
+      },
+      "model-00042-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506647820000000,
+        "ctime_ns": 1785506647820000000,
+        "device": 16777242,
+        "inode": 11084687
+      },
+      "model-00043-of-00048.safetensors": {
+        "bytes": 3568770544,
+        "mtime_ns": 1785506657140000000,
+        "ctime_ns": 1785506657140000000,
+        "device": 16777242,
+        "inode": 11098387
+      },
+      "model-00044-of-00048.safetensors": {
+        "bytes": 3590026352,
+        "mtime_ns": 1785506666700000000,
+        "ctime_ns": 1785506666700000000,
+        "device": 16777242,
+        "inode": 11112006
+      },
+      "model-00045-of-00048.safetensors": {
+        "bytes": 1059332516,
+        "mtime_ns": 1785506668200000000,
+        "ctime_ns": 1785506668200000000,
+        "device": 16777242,
+        "inode": 11125706
+      },
+      "model-00046-of-00048.safetensors": {
+        "bytes": 3610455184,
+        "mtime_ns": 1785506675660000000,
+        "ctime_ns": 1785506675660000000,
+        "device": 16777242,
+        "inode": 11129751
+      },
+      "model-00047-of-00048.safetensors": {
+        "bytes": 3560111960,
+        "mtime_ns": 1785506683250000000,
+        "ctime_ns": 1785506683250000000,
+        "device": 16777242,
+        "inode": 11143529
+      },
+      "model-00048-of-00048.safetensors": {
+        "bytes": 3692775244,
+        "mtime_ns": 1785506690860000000,
+        "ctime_ns": 1785506690860000000,
+        "device": 16777242,
+        "inode": 11157115
+      }
+    }
+  },
+  "profile": {
+    "path": "embedded-profile",
+    "source_profile_sha256": "6c25185500c07d4419516d30b0902e457c99bd70543c4f5bea026756d1804eb4",
+    "content": {
+      "name": "dsv4-ds4-quality128-native-mxfp4",
+      "engine": "ds4",
+      "description": "Quality-first 128 GB DS4 profile preserving exact native MXFP4 routed experts on ten sensitivity-selected layers, with IQ2_XXS/Q2_K elsewhere and native-MXFP4 DSpark down projections.",
+      "source": "apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8",
+      "output": "DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128",
+      "ds4_revision": "45e6cd76e849d86a8280e24ef43fa149ce9e330a",
+      "toolchain": {
+        "root": "antirez/ds4@ds4f-mxfp4",
+        "upstream_base_revision": "4893e0c40fba03dbc85555faeb035799aa04e0b6",
+        "quantizer_sha256": "f0a381f4ada808ea2afa740d964354fa327fc1235ba7cebf50874eb89fb97ac5",
+        "runtime_sha256": "2aaf20469b9918d6d6ab8787a02811c11228547cd979787879a03dba8a9e7824"
+      },
+      "main": {
+        "experts": "iq2_xxs",
+        "routed_w2": "q2_k",
+        "attention_proj": "q8_0",
+        "shared": "q8_0",
+        "preserve_indexer_f16": true,
+        "tensor_types": [
+          "output.weight=q8_0"
+        ],
+        "native_mxfp4_layers": [
+          10,
+          14,
+          30,
+          34,
+          37,
+          38,
+          39,
+          40,
+          41,
+          42
+        ],
+        "native_mxfp4_layer_selection": "Antirez 37-42 seed plus cross-recipe sensitivity controls 10,14,30,34; source packed I8 codes and F8_E8M0 scales are preserved exactly.",
+        "estimated_bytes": 102826238912
+      },
+      "dspark": {
+        "experts": "iq2_xxs",
+        "routed_w2": "mxfp4",
+        "block_size": 5,
+        "target_layers": [
+          40,
+          41,
+          42
+        ],
+        "markov_rank": 256,
+        "noise_token_id": 128799,
+        "calibration": "target-layer proxy aliases"
+      },
+      "verification": {
+        "require_imatrix_strict": true,
+        "require_native_byte_compare": true,
+        "main": {
+          "bytes": 102826238912,
+          "tensors": 1328,
+          "tensor_types": {
+            "i32": 3,
+            "iq2_xxs": 66,
+            "q2_k": 33,
+            "mxfp4": 30,
+            "f16": 359,
+            "f32": 492,
+            "q8_0": 345
+          }
+        },
+        "dspark": {
+          "bytes": 7297737120,
+          "tensors": 81,
+          "tensor_types": {
+            "f32": 34,
+            "f16": 7,
+            "q8_0": 31,
+            "iq2_xxs": 6,
+            "mxfp4": 3
+          }
+        }
+      },
+      "calibration": {
+        "repo_id": "ox-ox/DeepSeek-V4-Flash-0731-GGUF",
+        "revision": "6d58a3a36030c3ccb969bb5759fc6ae08cd299f8",
+        "filename": "imatrix/DeepSeek-V4-Flash-0731-chat-v2-routed-moe-ds4-1p5m.dat",
+        "bytes": 450892654,
+        "sha256": "6fce7674df701de544e5d3351aab04e67602eddeafeb48cf70e77ebe47239eb4",
+        "coverage_policy": "exact 129 routed names and vector dimensions; strict mode checks only imatrix-consuming output types"
+      },
+      "template": {
+        "role": "metadata-tokenizer-tensor-order-only",
+        "repo_id": "antirez/deepseek-v4-gguf",
+        "revision": "1cd7b564460821938add0475a60b942c409295e0",
+        "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf",
+        "remote_bytes": 86720111488,
+        "lfs_sha256": "ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0",
+        "xet_hash": "7da16e1025c1b856490c29c341f4e467d15cb195389c70383dade5e6108799ac",
+        "header_sha256": "f0e1d5e8f3b008402aa6eb32cada3873dd926c8bf5e7d00d7788eec65f09dd6d"
+      },
+      "recipe_control": {
+        "repo_id": "antirez/deepseek-v4-gguf",
+        "revision": "1cd7b564460821938add0475a60b942c409295e0",
+        "dspark_filename": "DeepSeek-V4-Flash-DSpark-support.gguf",
+        "weights_used": false
+      }
+    }
+  },
+  "calibration": {
+    "repo_id": "ox-ox/DeepSeek-V4-Flash-0731-GGUF",
+    "filename": "imatrix/DeepSeek-V4-Flash-0731-chat-v2-routed-moe-ds4-1p5m.dat",
+    "revision": "6d58a3a36030c3ccb969bb5759fc6ae08cd299f8",
+    "bytes": 450892654,
+    "sha256": "6fce7674df701de544e5d3351aab04e67602eddeafeb48cf70e77ebe47239eb4",
+    "coverage_validation": {
+      "entries": 129,
+      "scope": "all routed gate/up/down tensors",
+      "policy": "exact-name-and-vector-length",
+      "global_ds4_imatrix_strict": true
+    }
+  },
+  "template": {
+    "repo_id": "antirez/deepseek-v4-gguf",
+    "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf",
+    "revision": "1cd7b564460821938add0475a60b942c409295e0",
+    "role": "metadata-tokenizer-tensor-order-and-shapes-only",
+    "weights_used": false,
+    "remote_bytes": 86720111488,
+    "lfs_sha256": "ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0",
+    "xet_hash": "7da16e1025c1b856490c29c341f4e467d15cb195389c70383dade5e6108799ac",
+    "downloaded_header_bytes": 67108864,
+    "header_sha256": "f0e1d5e8f3b008402aa6eb32cada3873dd926c8bf5e7d00d7788eec65f09dd6d",
+    "resolved": {
+      "repo_id": "antirez/deepseek-v4-gguf",
+      "filename": "DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf",
+      "role": "metadata-tokenizer-tensor-order-and-shapes-only",
+      "weights_used": false,
+      "remote_bytes": 86720111488,
+      "lfs_sha256": "ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0",
+      "xet_hash": "7da16e1025c1b856490c29c341f4e467d15cb195389c70383dade5e6108799ac",
+      "resolved_etag": "7da16e1025c1b856490c29c341f4e467d15cb195389c70383dade5e6108799ac",
+      "range": "bytes=0-67108863",
+      "header_bytes": 67108864,
+      "header_sha256": "f0e1d5e8f3b008402aa6eb32cada3873dd926c8bf5e7d00d7788eec65f09dd6d"
+    }
+  },
+  "dspark_calibration_alias": {
+    "schema": "deepseek-model-tools.dspark-imatrix-alias.v1",
+    "function": "extend_dspark_legacy_imatrix",
+    "source": {
+      "repo_id": "ox-ox/DeepSeek-V4-Flash-0731-GGUF",
+      "revision": "6d58a3a36030c3ccb969bb5759fc6ae08cd299f8",
+      "filename": "imatrix/DeepSeek-V4-Flash-0731-chat-v2-routed-moe-ds4-1p5m.dat",
+      "bytes": 450892654,
+      "sha256": "6fce7674df701de544e5d3351aab04e67602eddeafeb48cf70e77ebe47239eb4"
+    },
+    "output_sha256": "689b446ed2e2657ebcb69a6516781ee0444e07641aeb6125233fb4e6fe7cbce3",
+    "mapping": [
+      {
+        "stage": 0,
+        "target_layer": 40,
+        "alias": "mtp.0.ffn_gate_exps.weight",
+        "source": "blk.40.ffn_gate_exps.weight"
+      },
+      {
+        "stage": 0,
+        "target_layer": 40,
+        "alias": "mtp.0.ffn_up_exps.weight",
+        "source": "blk.40.ffn_up_exps.weight"
+      },
+      {
+        "stage": 0,
+        "target_layer": 40,
+        "alias": "mtp.0.ffn_down_exps.weight",
+        "source": "blk.40.ffn_down_exps.weight"
+      },
+      {
+        "stage": 1,
+        "target_layer": 41,
+        "alias": "mtp.1.ffn_gate_exps.weight",
+        "source": "blk.41.ffn_gate_exps.weight"
+      },
+      {
+        "stage": 1,
+        "target_layer": 41,
+        "alias": "mtp.1.ffn_up_exps.weight",
+        "source": "blk.41.ffn_up_exps.weight"
+      },
+      {
+        "stage": 1,
+        "target_layer": 41,
+        "alias": "mtp.1.ffn_down_exps.weight",
+        "source": "blk.41.ffn_down_exps.weight"
+      },
+      {
+        "stage": 2,
+        "target_layer": 42,
+        "alias": "mtp.2.ffn_gate_exps.weight",
+        "source": "blk.42.ffn_gate_exps.weight"
+      },
+      {
+        "stage": 2,
+        "target_layer": 42,
+        "alias": "mtp.2.ffn_up_exps.weight",
+        "source": "blk.42.ffn_up_exps.weight"
+      },
+      {
+        "stage": 2,
+        "target_layer": 42,
+        "alias": "mtp.2.ffn_down_exps.weight",
+        "source": "blk.42.ffn_down_exps.weight"
+      }
+    ],
+    "tool": {
+      "repository_revision": "d97ba6b63ed857a5456f92a5190381fd7e45ca9f",
+      "importer_sha256": "b39987ce5ce35564bb4053ec11265c0ff41be233595ae17a6e039318a52dd4da"
+    },
+    "coverage_validation": {
+      "entries": 138,
+      "trunk_entries": 129,
+      "dspark_entries": 9,
+      "coverage": "exact-name-and-vector-length"
+    }
+  },
+  "tools": {
+    "ds4_revision": "d516d4eeb82c454aeb2831af1b1961801d6b571b",
+    "ds4_root": "antirez/ds4@ds4f-mxfp4",
+    "ds4_quantizer_sha256": "f0a381f4ada808ea2afa740d964354fa327fc1235ba7cebf50874eb89fb97ac5",
+    "ds4_runtime_sha256": "2aaf20469b9918d6d6ab8787a02811c11228547cd979787879a03dba8a9e7824",
+    "omlx_revision": "76352ed2363e42b2146243463875756e683a640d",
+    "deepseek_model_tools_revision": "e24746463a0a3e79036dd9c0472deac6ce704f08",
+    "deepseek_model_tools_worktree": {
+      "dirty": true,
+      "status_sha256": "59efc8a4b51d3dc33f267055db94edc9ac8e2392aa9f6e7033b3842a060b4ad2",
+      "tracked_patch_sha256": "658a58301aad033c5027f62ee1f1719ab0b01141c4b06b2d296f5f05c8f76aea"
+    },
+    "build_driver_sha256": "e9048f7535fe360756ce4505edae8f5d10f901e3fcc7759a37abbfbfc3949b4e"
+  },
+  "commands": [
+    [
+      "deepseek4-quantize",
+      "--hf",
+      "apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8",
+      "--template",
+      "antirez/deepseek-v4-gguf/ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0.header-only.sparse.gguf",
+      "--experts",
+      "iq2_xxs",
+      "--routed-w2",
+      "q2_k",
+      "--attention-proj",
+      "q8_0",
+      "--shared",
+      "q8_0",
+      "--threads",
+      "8",
+      "--imatrix",
+      "ox-ox/DeepSeek-V4-Flash-0731-GGUF/imatrix/DeepSeek-V4-Flash-0731-chat-v2-routed-moe-ds4-1p5m.dat",
+      "--tensor-type",
+      "output.weight=q8_0",
+      "--tensor-type",
+      "blk.2.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.4.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.6.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.8.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.10.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.12.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.14.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.16.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.18.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.20.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.22.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.24.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.26.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.28.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.30.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.32.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.34.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.36.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.38.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.40.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.42.indexer.attn_q_b.weight=f16",
+      "--tensor-type",
+      "blk.10.ffn_gate_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.10.ffn_up_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.10.ffn_down_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.14.ffn_gate_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.14.ffn_up_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.14.ffn_down_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.30.ffn_gate_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.30.ffn_up_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.30.ffn_down_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.34.ffn_gate_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.34.ffn_up_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.34.ffn_down_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.37.ffn_gate_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.37.ffn_up_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.37.ffn_down_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.38.ffn_gate_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.38.ffn_up_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.38.ffn_down_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.39.ffn_gate_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.39.ffn_up_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.39.ffn_down_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.40.ffn_gate_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.40.ffn_up_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.40.ffn_down_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.41.ffn_gate_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.41.ffn_up_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.41.ffn_down_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.42.ffn_gate_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.42.ffn_up_exps.weight=mxfp4",
+      "--tensor-type",
+      "blk.42.ffn_down_exps.weight=mxfp4",
+      "--imatrix-strict",
+      "--out",
+      "DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf"
+    ],
+    [
+      "deepseek4-quantize",
+      "--hf",
+      "apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8",
+      "--template",
+      "antirez/deepseek-v4-gguf/ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0.header-only.sparse.gguf",
+      "--dspark-support",
+      "--experts",
+      "iq2_xxs",
+      "--routed-w2",
+      "mxfp4",
+      "--attention-proj",
+      "q8_0",
+      "--shared",
+      "q8_0",
+      "--dspark-block-size",
+      "5",
+      "--dspark-markov-rank",
+      "256",
+      "--dspark-noise-token-id",
+      "128799",
+      "--dspark-target-layers",
+      "40,41,42",
+      "--threads",
+      "8",
+      "--imatrix",
+      "ox-ox/DeepSeek-V4-Flash-0731-GGUF/imatrix/DeepSeek-V4-Flash-0731-dspark-alias.dat",
+      "--out",
+      "DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128-DSpark-support.gguf",
+      "--imatrix-strict"
+    ]
+  ],
+  "appledouble_excluded": true
+}

BUILD_PLAN.md

@@ -0,0 +1,73 @@
+# Build and verification record
+
+## Objective
+
+Create a reproducible DS4 package that maximizes target-model quality within a
+128 GB unified-memory envelope. Preserve exact native MXFP4 routed experts on
+the ten sensitivity-selected layers and retain the proven low-bit/Q8 policy
+everywhere else.
+
+## Inputs and toolchain
+
+- All weights were regenerated from
+  `apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8`.
+- The sparse reference GGUF supplied metadata, tokenizer, tensor order and
+  shapes only.
+- All 48 source-shard hashes, source metadata and the abliteration manifest
+  were verified before conversion.
+- The routed-expert imatrix supplied exact coverage for the 129 tensors whose
+  output types consume calibration data.
+- DS4 source revision: `d516d4eeb82c454aeb2831af1b1961801d6b571b`.
+- Upstream `ds4f-mxfp4` base: `4893e0c40fba03dbc85555faeb035799aa04e0b6`.
+- Quantizer SHA-256: `f0a381f4ada808ea2afa740d964354fa327fc1235ba7cebf50874eb89fb97ac5`.
+- Runtime SHA-256: `2aaf20469b9918d6d6ab8787a02811c11228547cd979787879a03dba8a9e7824`.
+
+## Quantization policy
+
+- Native MXFP4 gate/up/down routed experts on layers
+  `10, 14, 30, 34, 37, 38, 39, 40, 41, 42`.
+- IQ2_XXS gate/up and Q2_K down routed experts on the other 33 MoE layers.
+- Q8 attention, shared-expert and output tensors.
+- Protected F16 indexer and auxiliary tensors.
+- DSpark support with IQ2_XXS gate/up and native MXFP4 down projections for
+  target layers 40, 41 and 42.
+
+## Conversion gates
+
+- Strict imatrix mode passed for main and DSpark conversions.
+- Main GGUF contains exactly 1,328 tensors and the intended 30 MXFP4 tensors.
+- DSpark support contains exactly 81 tensors and three MXFP4 down aggregates.
+- Every native tensor comparison required an explicit `byte_compare: OK`.
+- All 30 main and three DSpark MXFP4 tensors independently reproduced the
+  source FP4 codes and scale bytes.
+- `ds4 --cpu --inspect --dspark-strict` reported zero missing tensors, invalid
+  bindings or metadata errors.
+- The final GGUF hashes are recorded in `SHA256SUMS`.
+
+## Result
+
+| Component | Bytes | GiB |
+| --- | ---: | ---: |
+| Main GGUF | `102,826,238,912` | `95.7644` |
+| DSpark support | `7,297,737,120` | `6.7965` |
+| Combined | `110,123,976,032` | `102.5609` |
+
+Native MXFP4 improves fidelity over a second Q4_K requantization and uses 4.25
+bits per weight rather than 4.50. Native MXFP4 Metal and the mixed DSpark path
+are newer than mature Q4_K kernels, so throughput should be measured on the
+target system.
+
+## One-million-token context
+
+Recommended target-only mode:
+
+```sh
+ds4 --metal \
+  -m DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf \
+  --ctx 1048576 --prefill-chunk 2048
+```
+
+Estimated residency is at most `110.30 GiB`. DSpark can be tested with a 1,024
+token prefill chunk; estimated residency is `114.02 GiB`. A 4,096-token chunk
+with DSpark is estimated at `123.24 GiB` and is not a reliably resident mode on
+a machine whose Metal recommended working set is approximately `121.60 GiB`.

DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128-DSpark-support.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd8593a232c9feebc4c91855d5ab486b17250fc8bc2f294bc80401f93b371566
+size 7297737120

DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2cfc36b761b59ea43531e7cdb02a690436a330e42ad57cb162726b385914df59
+size 102826238912

PROVENANCE.md

@@ -0,0 +1,22 @@
+# Provenance
+
+This artifact was regenerated from the abliterated FP8 checkpoint only. No
+model tensors were copied from a public GGUF or another quantized model.
+
+- Weight source: `apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8`
+- Upstream base model: `deepseek-ai/DeepSeek-V4-Flash-0731`
+- Upstream revision: `9e165c30e2704aec5d9d593cce3eebd58bbef1cb`
+- Abliteration manifest SHA-256: `8e6f40f3d92542720aabdafb728f90e3397fba68c6a8d7ba91772b5dd41bc767`
+- Calibration: `ox-ox/DeepSeek-V4-Flash-0731-GGUF/imatrix/DeepSeek-V4-Flash-0731-chat-v2-routed-moe-ds4-1p5m.dat`
+- Calibration SHA-256: `6fce7674df701de544e5d3351aab04e67602eddeafeb48cf70e77ebe47239eb4`
+- DSpark calibration proxy SHA-256: `689b446ed2e2657ebcb69a6516781ee0444e07641aeb6125233fb4e6fe7cbce3`
+- Metadata template: `antirez/deepseek-v4-gguf/DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf`
+- Template LFS SHA-256: `ca22ae2f838e14077c22bc1c1417b71b45b5e5a3687bd96c2ac6e17fdb6261c0`
+- Template Xet object: `7da16e1025c1b856490c29c341f4e467d15cb195389c70383dade5e6108799ac`
+- Template 64 MiB header SHA-256: `f0e1d5e8f3b008402aa6eb32cada3873dd926c8bf5e7d00d7788eec65f09dd6d`
+- Engine: `ds4`
+- DSpark support: yes
+
+The source abliteration modified 36 attention `wo_b` tensors. It did not modify
+routed-expert codes or scales, allowing the selected MXFP4 experts to be
+repacked losslessly from their original I8 codes and F8_E8M0 scales.

README.md

@@ -0,0 +1,128 @@
+# DeepSeek V4 Flash Abliterated — DS4 Quality128
+
+> **Exact MXFP4 experts, maximum resident quality.**
+
+> **Runtime requirement:** until native MXFP4 support is merged into DS4's
+> default branch, use the
+> official [`antirez/ds4` `ds4f-mxfp4` branch](https://github.com/antirez/ds4/tree/ds4f-mxfp4).
+
+This is a quality-first DS4 package designed to keep DeepSeek V4 Flash resident
+on a 128 GB M1 Ultra while preserving the most sensitive routed experts in
+their exact native MXFP4 representation.
+
+All weights were regenerated from the abliterated FP8 checkpoint
+[`apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8`](https://huggingface.co/apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8).
+
+No model tensor was copied from another GGUF. The source abliteration edited
+only 36 attention `wo_b` tensors; routed-expert codes and scales were untouched.
+
+## Artifacts
+
+| File | Bytes | GiB | SHA-256 |
+| --- | ---: | ---: | --- |
+| `DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf` | `102,826,238,912` | `95.7644` | `2cfc36b761b59ea43531e7cdb02a690436a330e42ad57cb162726b385914df59` |
+| `DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128-DSpark-support.gguf` | `7,297,737,120` | `6.7965` | `cd8593a232c9feebc4c91855d5ab486b17250fc8bc2f294bc80401f93b371566` |
+| **Combined** | **`110,123,976,032`** | **`102.5609`** | — |
+
+## Quantization profile
+
+- Exact native MXFP4 gate/up/down routed experts on layers
+  `10, 14, 30, 34, 37, 38, 39, 40, 41, 42`.
+- IQ2_XXS gate/up and Q2_K down routed experts on the other 33 MoE layers.
+- Q8 attention, shared-expert and output paths.
+- F16 protected indexer and auxiliary tensors.
+- The routed-expert imatrix applies only to genuinely requantized IQ2/Q2
+  tensors; preserved MXFP4 needs no imatrix.
+- DSpark support uses IQ2_XXS gate/up and exact native MXFP4 down projections
+  for target layers 40, 41 and 42.
+
+Main GGUF type histogram (1,328 tensors): F32 492, F16 359, I32 3, Q8_0 345,
+IQ2_XXS 66, Q2_K 33 and MXFP4 30. DSpark support histogram (81 tensors): F32
+34, F16 7, Q8_0 31, IQ2_XXS 6 and MXFP4 3.
+
+## Runtime requirement
+
+Native MXFP4 and this DSpark layout require the official
+[`antirez/ds4` `ds4f-mxfp4` branch](https://github.com/antirez/ds4/tree/ds4f-mxfp4),
+or DS4's default branch after those changes are merged. The artifact was
+qualified against `ds4f-mxfp4` revision
+`4893e0c40fba03dbc85555faeb035799aa04e0b6` with the DSpark generation fix
+tracked in [antirez/ds4#642](https://github.com/antirez/ds4/issues/642).
+
+- Quantizer SHA-256: `f0a381f4ada808ea2afa740d964354fa327fc1235ba7cebf50874eb89fb97ac5`
+- Runtime SHA-256: `2aaf20469b9918d6d6ab8787a02811c11228547cd979787879a03dba8a9e7824`
+
+Do not assume an older DS4 executable can load or execute this package
+correctly merely because it accepts the GGUF container.
+
+## Launch examples
+
+Normal target-only inference:
+
+```sh
+ds4 --metal \
+  -m DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf
+```
+
+Greedy DSpark inference:
+
+```sh
+ds4 --metal \
+  -m DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf \
+  --mtp DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128-DSpark-support.gguf \
+  --dspark --temp 0
+```
+
+## One-million-token context
+
+`--ctx 1048576` counts prompt and completion together. The safest maximum-quality
+resident mode omits DSpark and uses a 2,048-token prefill chunk:
+
+```sh
+ds4 --metal \
+  -m DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf \
+  --ctx 1048576 --prefill-chunk 2048
+```
+
+Estimated residency is at most `110.30 GiB`, leaving at least `11.30 GiB`
+below Metal's approximately `121.60 GiB` recommended working set. Omitting
+DSpark does not reduce target-model quality; it only forgoes speculative decode.
+
+After measuring peak memory on this machine, DSpark can be enabled with the
+smaller chunk:
+
+```sh
+ds4 --metal \
+  -m DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf \
+  --ctx 1048576 --prefill-chunk 1024 \
+  --mtp DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128-DSpark-support.gguf \
+  --dspark --temp 0
+```
+
+That mode is estimated at `114.02 GiB`, about `7.58 GiB` below the recommended
+working set. DSpark with chunk 4096 is estimated at `123.24 GiB` and is not a
+reliably resident configuration.
+
+## Verification and provenance
+
+The build passed strict main and DSpark planning, exact size/type/name-set
+checks, strict imatrix coverage, source validation, and byte reproduction for
+all 30 main plus three DSpark MXFP4 tensors. `SHA256SUMS` binds the final GGUFs
+and documentation. See:
+
+- [`BUILD_MANIFEST.json`](BUILD_MANIFEST.json) for pinned source shards, tools,
+  commands and publication metadata.
+- [`PROVENANCE.md`](PROVENANCE.md) for the compact lineage record.
+- [`BUILD_PLAN.md`](BUILD_PLAN.md) for the completed build gates and benchmark
+  matrix.
+- [`ds4-upstream-issues.md`](ds4-upstream-issues.md) for remaining DS4 runtime
+  and memory improvements.
+
+The sibling MLX package
+`DeepSeek-V4-Flash-0731-Abliterated-MLX-Quality128-suboptimal` is intentionally
+retained for speed comparison; it is not the canonical quality artifact.
+
+Native MXFP4 Metal kernels and the mixed DSpark path are newer than the mature
+Q4_K path. Correctness and throughput should therefore be benchmarked against
+the archived DS4 v1 and retained MLX comparator before changing everyday launch
+defaults.

SHA256SUMS

@@ -0,0 +1,7 @@
+27e3fbe612341ffe8f0060014831d88b51b2f1ca986f820d98296e4e6b7c85a7  BUILD_MANIFEST.json
+b52972dd4cf914a8abe0c224b39ac5975c808450595e87945bc7a00c37eac649  BUILD_PLAN.md
+cd8593a232c9feebc4c91855d5ab486b17250fc8bc2f294bc80401f93b371566  DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128-DSpark-support.gguf
+2cfc36b761b59ea43531e7cdb02a690436a330e42ad57cb162726b385914df59  DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf
+ea2f1f2bdb10266e509dbe376a5f5b51eba6433ebecc43c8a0d1b3c87d2aaa02  PROVENANCE.md
+6b8c9bc266d71ac941ded6bd3750d7716f4158507446b8ac42a517b9872a9049  README.md
+afd8c0883b69a9280ccbd6e191d6f9abfe757498e1cfbd1c9f949edf4ef1a95b  ds4-upstream-issues.md

ds4-upstream-issues.md

@@ -0,0 +1,163 @@
+# DS4 upstream issues and improvement candidates
+
+Audit base: the official
+[`antirez/ds4` `ds4f-mxfp4` branch](https://github.com/antirez/ds4/tree/ds4f-mxfp4)
+at revision `4893e0c40fba03dbc85555faeb035799aa04e0b6`.
+
+This is an engineering backlog, not a claim that issues have already been filed upstream. Items distinguish blockers for the planned artifact from optional improvements to DS4 itself.
+
+## P0 — required for the Quality128 package
+
+### 1. DSpark planner rejects preserved MXFP4 routed tensors
+
+**Status:** fixed on the local PR branch and tracked upstream as
+[antirez/ds4#642](https://github.com/antirez/ds4/issues/642). The real strict
+support dry-run now reports 81 tensors, three MXFP4 down aggregates, and
+`7,297,737,120` bytes.
+
+**Observed:**
+
+```text
+--dspark-support --routed-w2 mxfp4 --dry-run
+error: unsupported DSpark planned tensor type
+```
+
+The main planner recognizes a narrowly scoped preserved-MXFP4 expert exception around `gguf-tools/deepseek4-quantize.c:1734`. The DSpark sizing path at approximately `gguf-tools/deepseek4-quantize.c:2232` accepts I32 or ordinary quantizable targets but does not mirror that exception.
+
+The DSpark generator already routes `mtp.*` expert tensors through the common native repacker. This appears to be a planning omission, not a missing conversion implementation.
+
+**Proposed change:** permit MXFP4 only when the DSpark plan entry is a packed routed expert with the required block alignment and source dtype/shape contract. Do not make MXFP4 a generic quantizable type.
+
+**Acceptance:**
+
+- Strict DSpark dry-run succeeds with `routed_w2=mxfp4`.
+- Exactly three DSpark down aggregates become MXFP4.
+- Expected support size is `7,297,737,120` bytes unless reviewed metadata changes explain otherwise.
+- Non-expert or incompatible source tensors are still rejected.
+
+### 2. Missing mixed IQ2/IQ2/MXFP4 DSpark Metal coverage
+
+**Status:** covered on the local PR branch by the hardened mixed-path Metal
+fixture for token counts one through five and the native-C quantizer regression
+fixture. Full support-model runtime smoke remains an artifact acceptance gate.
+
+Current MXFP4 Metal coverage tests all-MXFP4 gate/up/down at limited batch sizes. The planned sidecar uses IQ2_XXS gate/up and MXFP4 down with DSpark proposal blocks of one through five tokens.
+
+**Proposed test matrix:**
+
+- Token counts `1, 2, 3, 4, 5`.
+- IQ2_XXS fused gate/up plus MXFP4 down.
+- Fused sum path for one through four tokens.
+- Generic MV plus expert-summation path for five tokens.
+- CPU/scalar reference comparison with explicit tolerances.
+- At least one full DSpark verifier smoke test using the real support binding.
+
+**Acceptance:** deterministic pass on M1 Ultra Metal with no out-of-bounds access and bounded numerical error versus the reference.
+
+### 3. Byte-compare failure should fail the process
+
+**Status:** still an upstream hardening opportunity. The Quality128 build
+wrapper closes the production risk by requiring the literal
+`byte_compare: OK` for every one of the 30 main and three DSpark native tensors.
+
+`gguf-tools/deepseek4-quantize.c` prints `byte_compare: FAIL` but the comparison path can still exit successfully. A production wrapper can parse for an explicit `byte_compare: OK`, but the quantizer should provide reliable process semantics itself.
+
+**Proposed change:** return nonzero when any byte mismatch is found or when a requested comparison cannot be completed.
+
+**Acceptance:** a deliberately corrupted reference tensor produces a nonzero exit; an exact native repack produces zero and `byte_compare: OK`.
+
+## P1 — high-value Metal/runtime improvements
+
+### 4. Right-size the one-million-context `comp_mask` allocation
+
+At `ctx=1,048,576` and prefill chunk `4096`, both `indexer_scores_by_tier` and `comp_mask_by_tier` are allocated as `comp_cap * prefill_cap * sizeof(float)` near `ds4.c:17158`.
+
+Each allocation is roughly `4 GiB`. The normal indexed batch path consumes compact top-k selected indices, while `comp_mask` appears to be used principally by one-row/fallback attention and small top-value scratch paths. If confirmed across all backends, allocating a full context-by-prefill mask is unnecessary.
+
+**Potential improvement:** size `comp_mask` to its maximum actual live row count, or give the large batch path a separate compact scratch allocation.
+
+**Expected benefit:** reclaim close to `4 GiB` at 1M/chunk4096, potentially moving the full resident Quality128 package plus DSpark from approximately `123.24` to `119.24 GiB` without changing weights.
+
+**Required proof:** audit every Metal/CUDA use, add allocation-bound assertions, run indexed-prefill/decode consistency tests, and demonstrate bit-identical or tolerance-identical logits before and after the change.
+
+### 5. Tile indexed-attention scoring independently of the MoE prefill chunk
+
+`indexer_scores` genuinely scales with compressed-cache rows times the global prefill batch. Reducing `--prefill-chunk` from 4096 to 1024 cuts total planned memory from about `123.24` to `114.02 GiB`, but also makes every prefill stage use smaller batches.
+
+**Potential improvement:** keep a larger MoE/dense prefill batch while processing indexer score/top-k work in bounded subtiles. This separates the attention-scratch memory decision from the MoE throughput decision.
+
+**Acceptance:** materially lower peak memory at 1M context, no retrieval/logit regression, and better prefill throughput than globally reducing the chunk to the same scratch footprint.
+
+### 6. Phase-separate prefill workspace and DSpark residency
+
+DSpark does not accelerate prefill, yet its approximately `6.80 GiB` support model and verifier state coexist with the largest prefill workspace. The prefill workspace is also retained after prefill for possible later session work.
+
+**Potential improvement:**
+
+1. Prefill before mapping/requesting residency for the DSpark support model.
+2. Release or shrink prefill-only tensors when entering decode.
+3. Map DSpark and allocate verifier state for speculative decode.
+4. Reverse the transition safely if a later prompt extension needs batched prefill.
+
+This is a lifecycle/allocator change, not a model-quality trade-off.
+
+**Acceptance:** chunk4096 prefill and DSpark decode both remain under the Metal working-set limit in their respective phases; session extension, cancellation and cleanup tests show no leaks or stale views.
+
+### 7. Startup memory reporting omits important persistent workspace
+
+The high-level planned-memory log accounts for model span, KV and its `scratch_bytes` estimate, but does not make all batch-prefill tensors, DSpark verifier snapshots, logits/host buffers and driver transients obvious. At 1M this can understate the actionable footprint by multiple GiB.
+
+**Potential improvement:** have the session allocator report:
+
+- Model and support-model mapped/resident bytes.
+- Raw and compressed KV.
+- Indexed-attention scratch.
+- Other prefill workspace.
+- Decode and DSpark verifier workspace.
+- Runtime tensor live/peak totals.
+- Metal recommended working set and remaining margin.
+
+The same complete calculation should gate allocation before a predictable out-of-memory failure.
+
+### 8. SSD streaming is incompatible with `--mtp`, and mixed-size caching is weak
+
+Current DS4 rejects `--ssd-streaming` with `--mtp` around `ds4.c:56160`. Separately, the streaming cache uses the dominant routed-expert size class. In this hybrid, the 33 IQ2/Q2 layers define that slab while the ten larger MXFP4 layers bypass it through mapped-model views.
+
+**Potential improvements:**
+
+- Make support-model residency compatible with main-model expert streaming.
+- Account for DSpark and long-context buffers when choosing an automatic cache budget.
+- Support two expert-cache size classes or a byte-addressed allocator so native-MXFP4 layers can participate.
+
+**Acceptance:** explicit memory caps are respected, native expert bytes remain exact, and cache-miss/decode benchmarks show predictable behavior on an external SSD.
+
+## P2 — build and release hardening
+
+### 9. Include the native Metal MXFP4 test in the default relevant test suite
+
+The scalar MXFP4 test is covered by the general test target, while `test-mxfp4-metal` is separate. A source revision can therefore pass routine tests without exercising the new native Metal kernel.
+
+**Proposed change:** add a documented aggregate target suitable for Metal release qualification, or include the test when Metal is available.
+
+### 10. Make runtime Metal-source provenance explicit
+
+DS4 compiles/loads Metal sources dynamically and can depend on the working directory, `--chdir`, or `DS4_METAL_MOE_SOURCE`. A current executable paired with stale or missing Metal source is a reproducibility risk.
+
+**Potential improvement:** embed a Metal-source revision/hash in the executable and print the resolved source path plus hash at startup. Release verification should fail on a mismatch when native MXFP4 is required.
+
+### 11. Expose machine-readable plan and inspection output
+
+Dry-run and inspection text is currently parsed by external tooling. A stable JSON mode would make it safer to assert tensor counts, types, byte sizes, imatrix coverage, support binding and memory planning.
+
+**Acceptance:** JSON schema/version is explicit and text output remains available for humans.
+
+## Suggested upstream submission order
+
+1. DSpark preserved-MXFP4 planner fix plus mixed-path test.
+2. Byte-compare exit semantics.
+3. `comp_mask` allocation audit and minimal right-sizing patch.
+4. Complete memory-reporting/accounting patch.
+5. Tiled indexer scratch and phase-separated residency as separately benchmarked changes.
+6. Streaming/dual-size-cache work only after resident Quality128 is stable.
+
+Each submission should be small, independently testable and accompanied by before/after memory or throughput measurements rather than bundled into the model-build patch.