@bendyline/gezel-inference 1.0.7 → 1.0.9

package/dist/catalog/chat-models/or/ornith1.5-9b-q4/manifest.json

@@ -136,6 +136,11 @@
           "enableThinking": false
         }
       }
+    },
+    "taskStrategy": {
+      "strategy": "outcome",
+      "procedure": "none",
+      "basis": "Strategy screen 2026-11-04 (6 books x 2, same build as its Qwen 3.5 9B base): outcome without the recipe passed 11 of 12 against stepwise 6 of 12, judged quality +0.69 [+0.03, +1.30] at 0.16x the wall time. The recipe added nothing (11 of 12, judged 4.4 against 4.7)."
     }
   },
   "llamaCpp": {

package/dist/catalog/chat-models/or/ornith1.5-9b-q8/manifest.json

@@ -136,6 +136,11 @@
           "enableThinking": false
         }
       }
+    },
+    "taskStrategy": {
+      "strategy": "outcome",
+      "procedure": "none",
+      "basis": "Inherited from ornith1.5-9b-q4, the same weights at Q4, not measured at Q8. That screen (2026-11-04, 6 books x 2): outcome without the recipe passed 11 of 12 against stepwise 6 of 12 at 0.16x the wall time."
     }
   },
   "llamaCpp": {

package/dist/catalog/chat-models/qw/qwen3.5-9b-q4/manifest.json

@@ -126,6 +126,11 @@
           "nMax": 4
         }
       }
+    },
+    "taskStrategy": {
+      "strategy": "outcome",
+      "procedure": "none",
+      "basis": "Round 2 (2026-11-02) and a screen (2026-11-04): outcome beat stepwise at 0.3-0.4x the wall time. The recipe overflowed a 32K window on powerpoint-deck (0 of 3). Without it the 9B ties the recipe on five books; on powerpoint-deck it passed 3 of 5, losing two to DocBlocks save_artifact rejecting ifExists replace on a new file with no hint."
     }
   },
   "ollama": {

package/dist/catalog/chat-models/qw/qwen3.8-27b-q4/manifest.json

@@ -176,6 +176,11 @@
           "nMax": 4
         }
       }
+    },
+    "taskStrategy": {
+      "strategy": "outcome",
+      "procedure": "none",
+      "basis": "Strategy benchmark rounds 1 and 2 (90 pairs): with and without the recipe score the same (judged quality +0.04 [-0.13, +0.19]); the recipe costs 1.37x the wall time, and round 1's quality edge did not repeat."
     }
   },
   "llamaCpp": {

package/dist/catalog/chat-models/qw/qwen3.8-flash-next-q2/manifest.json

@@ -3,7 +3,7 @@
   "kind": "chat-model",
   "id": "qwen3.8-flash-next-q2",
   "name": "Qwen 3.8 Flash Next (180B-A6B, IQ2_XXS/Q2_K, DS4)",
-  "description": "An experimental preview of the architecture behind Qwen4: a 125B-parameter MoE with 6B active per token, plus a 51B n-gram embedding table. This antirez DS4 Q2 build is a single 137.10 GiB GGUF with 41.73 GiB of resident main/MTP weights; its original 95.37 GiB BF16 n-gram table remains on a fast local SSD and is read row-by-row without being mapped or preloaded. It is the practical DS4 variant for Apple-silicon Macs, starting at a smaller context on memory-constrained systems.",
+  "description": "An experimental preview of the architecture behind Qwen4: a 125B-parameter MoE with 6B active per token, plus a 51B n-gram embedding table. This antirez DS4 Q2 build is a single 137.10 GiB GGUF with 41.73 GiB of resident main/MTP weights; its original 95.37 GiB BF16 n-gram table remains on a fast local SSD and is read row-by-row without being mapped or preloaded. It is the practical DS4 variant for Apple-silicon Macs: about 47 GiB fully loaded at 128K context on 96 GiB+ machines. On Macs with less memory, DS4 streams its routed experts from SSD with an 8 GiB expert cache, about 22 GiB in all at 64K context, at a lower speed.",
   "tags": [
     "qwen",
     "agentic",
@@ -28,8 +28,8 @@
     "name": "Qwen",
     "url": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next"
   },
-  "version": "1.0.2",
-  "updatedAt": "2026-09-17T00:00:00Z",
+  "version": "1.0.3",
+  "updatedAt": "2026-11-05T00:00:00Z",
   "license": "qwen-community-1.0",
   "licenseClass": "custom-restricted",
   "licenseShortName": "Qwen Community 1.0",
@@ -186,15 +186,32 @@
     "quantization": "IQ2_XXS/Q2_K",
     "visionEncoder": {
       "huggingfaceRepo": "ggml-org/Qwen3.8-Flash-Next-GGUF",
-      "revision": "01534bc2e1877d5de995b73d247d4459d273e688",
+      "revision": "91fb00353e766eb43944ed223665ea4fb87dad47",
       "filename": "mmproj-Qwen3.8-Flash-Next-Q8_0.gguf",
       "sha256": "b2e9b5e4a44c107f8867e67dbf09b607fd99ae33c1a97a60a6720aeb252a9dad",
       "sizeBytes": 616703104
     },
-    "residentBytes": 55834574848,
+    "residentBytes": 23053236961,
     "residentWeightBytes": 44807246316,
-    "ssdStreaming": false,
-    "ssdStreamingSupported": false,
+    "kvBytesPerToken": 72581,
+    "residentCtxTokens": 65536,
+    "cacheExpertsBytes": 8589934592,
+    "ssdStreaming": true,
+    "ssdStreamingSupported": true,
+    "ssdStreamingBackends": [
+      "metal"
+    ],
+    "fullResidencyMemory": {
+      "workingSetBytes": 48640504627,
+      "ctxTokens": 65536,
+      "bytesPerToken": 36262
+    },
+    "streamingMemory": {
+      "fixedBytes": 14463302369,
+      "ctxTokens": 65536,
+      "bytesPerToken": 72581,
+      "minCacheBytes": 4294967296
+    },
     "prefillChunk": 1024,
     "mtp": {
       "exactSampling": true

package/dist/catalog/chat-models/qw/qwen3.8-flash-next-q2/versions/1.0.3/manifest.json

@@ -0,0 +1,47 @@
+{
+  "schemaVersion": 1,
+  "version": "1.0.3",
+  "releasedAt": "2026-11-05T00:00:00Z",
+  "minGezelVersion": "1.26309",
+  "approxSizeBytes": 147823830144,
+  "ds4": {
+    "huggingfaceRepo": "antirez/qwen3.8-flash-next-gguf",
+    "revision": "d600fe1a43d2e1cdcadb85144ce3142f66f9eefe",
+    "filename": "Qwen3.8-Flash-Next-Q2.gguf",
+    "sha256": "b1b93fa69aca5f187b0fb813aca8f3ec1beb5cf8cf0bd38cf041b93e0b6ccac9",
+    "approxSizeBytes": 147207127040,
+    "quantization": "IQ2_XXS/Q2_K",
+    "visionEncoder": {
+      "huggingfaceRepo": "ggml-org/Qwen3.8-Flash-Next-GGUF",
+      "revision": "91fb00353e766eb43944ed223665ea4fb87dad47",
+      "filename": "mmproj-Qwen3.8-Flash-Next-Q8_0.gguf",
+      "sha256": "b2e9b5e4a44c107f8867e67dbf09b607fd99ae33c1a97a60a6720aeb252a9dad",
+      "sizeBytes": 616703104
+    },
+    "residentBytes": 23053236961,
+    "residentWeightBytes": 44807246316,
+    "kvBytesPerToken": 72581,
+    "residentCtxTokens": 65536,
+    "cacheExpertsBytes": 8589934592,
+    "ssdStreaming": true,
+    "ssdStreamingSupported": true,
+    "ssdStreamingBackends": [
+      "metal"
+    ],
+    "fullResidencyMemory": {
+      "workingSetBytes": 48640504627,
+      "ctxTokens": 65536,
+      "bytesPerToken": 36262
+    },
+    "streamingMemory": {
+      "fixedBytes": 14463302369,
+      "ctxTokens": 65536,
+      "bytesPerToken": 72581,
+      "minCacheBytes": 4294967296
+    },
+    "prefillChunk": 1024,
+    "mtp": {
+      "exactSampling": true
+    }
+  }
+}