Quantized MTP draft head (INT4/g64 affine, all 8 head matrices)

mtp.safetensors: 239 MB (was 849 MB). Trunk weights are
unchanged. Acceptance re-verified flat-or-better against the
previous head before publishing; mtplx_runtime.json carries the
fresh fingerprint-bound verification rows and config.json declares
the prequantized head layout (loadable by MTPLX >= 2.0.1;
this model family needs >= 2.7.0).

config.json

@@ -912,5 +912,13 @@
     "temporal_patch_size": 2
   },
   "vision_end_token_id": 248054,
-  "vision_start_token_id": 248053
+  "vision_start_token_id": 248053,
+  "mtplx_mtp_quantization": {
+    "bits": 4,
+    "group_size": 64,
+    "mode": "affine",
+    "policy": "all",
+    "prequantized": true,
+    "description": "All 8 MTP draft-head matrices (fc + attention q/k/v/o + MLP gate/up/down) packed MLX INT4/g64 affine from the released sidecar; head norms keep the pack's float dtype. Verified flat-or-better acceptance vs the unquantized head before publishing."
+  }
 }

mtp.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6ee6f23703190a9468baab968205b4256cee628cfcf32cd79e8592524a8fdbc7
-size 849400388
+oid sha256:f803737b0e64bcce54fcfa22f2c7ee342ad37192b8e2a3299fc940387a828c9d
+size 238934226

mtplx_runtime.json

@@ -91,7 +91,18 @@
     "recommended_profile_note": "Restamped sustained->turbo 2026-08-14 23:50 (mirrors the OQ restamp): the sustained recommendation came from the forge-verify rows; every gated dropday arm, the CLI turbo-default resolution for this public id, and the installed-app receipt (55.5 tok/s, draft 1.0) all served turbo. Served behavior unchanged: under turbo the hidden profile-scoped contract fell back to draft=target sampler (1.0/0.95/20), identical to the recommended_draft_sampler stamp now visible.",
     "source_format": "bf16_native",
     "source_repo": "/Users/youssof/.mtplx/models/Qwen--Qwen3.8-27B",
-    "source_sha": null
+    "source_sha": null,
+    "head_quantization": {
+      "bits": 4,
+      "group_size": 64,
+      "mode": "affine",
+      "policy": "all",
+      "quantized_at": "2026-08-20T04:49:57Z",
+      "source_sidecar_bytes": 849400388,
+      "quantized_sidecar_bytes": 238934226,
+      "tool": "scripts/build_qwen38_q4head_sidecar.py",
+      "note": "Structural head quantization of the released sidecar; no calibration, no training. Trunk weights unchanged."
+    }
   },
   "mtp_contract": {
     "base_hidden_variant": "post_norm",
@@ -103,8 +114,8 @@
   },
   "mtp_depth_default": 3,
   "mtp_depth_max": 3,
-  "mtp_sidecar": "bf16",
-  "mtplx_version": "2.6.0",
+  "mtp_sidecar": "int4-g64-prequantized",
+  "mtplx_version": "2.9.0",
   "precision_policy": {
     "intended_default_for": [
       "m1",
@@ -144,14 +155,18 @@
     "top_p": 0.95
   },
   "speed_evidence": {
+    "depth": 3,
+    "tok_s": [
+      45.38375130593394
+    ],
     "acceptance_by_depth": [
-      0.9790209790209791,
-      0.8671328671328671,
-      0.7762237762237763
+      0.9384057971014492,
+      0.855072463768116,
+      0.7427536231884058
     ],
-    "acceptance_collapsed": [],
-    "depth": 3,
-    "failure_reasons": [],
+    "greedy_diagnostic": {
+      "tok_s": 19.60451265560769
+    },
     "forge_verify_rows": [
       {
         "acceptance_by_position": [],
@@ -163,12 +178,12 @@
         "hit_token_budget_count": 0,
         "multiplier_vs_ar": 1.0,
         "quality_passed": true,
-        "tok_s": 20.991412027515594,
-        "verify_time_s": 56.19218879099935
+        "tok_s": 19.60451265560769,
+        "verify_time_s": 30.370817083050497
       },
       {
         "acceptance_by_position": [
-          0.9635416666666666
+          0.9712460063897763
         ],
         "depth": 1,
         "finish_reasons": {
@@ -176,15 +191,15 @@
         },
         "hit_token_budget": false,
         "hit_token_budget_count": 0,
-        "multiplier_vs_ar": 1.848048265782575,
+        "multiplier_vs_ar": 1.5051947774660717,
         "quality_passed": true,
-        "tok_s": 38.79314259377768,
-        "verify_time_s": 21.435596173687372
+        "tok_s": 29.5086100639882,
+        "verify_time_s": 20.34781050006859
       },
       {
         "acceptance_by_position": [
-          0.9818731117824774,
-          0.918429003021148
+          0.9889624724061811,
+          0.9448123620309051
         ],
         "depth": 2,
         "finish_reasons": {
@@ -192,16 +207,16 @@
         },
         "hit_token_budget": false,
         "hit_token_budget_count": 0,
-        "multiplier_vs_ar": 1.210236493022171,
+        "multiplier_vs_ar": 1.976632257031939,
         "quality_passed": true,
-        "tok_s": 25.40457287576389,
-        "verify_time_s": 38.888131326646544
+        "tok_s": 38.75091209846504,
+        "verify_time_s": 34.64071964554023
       },
       {
         "acceptance_by_position": [
-          0.9790209790209791,
-          0.8671328671328671,
-          0.7762237762237763
+          0.9384057971014492,
+          0.855072463768116,
+          0.7427536231884058
         ],
         "depth": 3,
         "finish_reasons": {
@@ -209,26 +224,23 @@
         },
         "hit_token_budget": false,
         "hit_token_budget_count": 0,
-        "multiplier_vs_ar": 2.4180184402971334,
+        "multiplier_vs_ar": 2.3149645238925305,
         "quality_passed": true,
-        "tok_s": 50.75762137040775,
-        "verify_time_s": 8.885614896658808
+        "tok_s": 45.38375130593394,
+        "verify_time_s": 21.260114512755536
       }
     ],
-    "greedy_diagnostic": {
-      "tok_s": 20.991412027515594
-    },
     "quality_rejected": [],
-    "tok_s": [
-      50.75762137040775
-    ],
-    "verdict": "mtp_depth_wins"
+    "acceptance_collapsed": [],
+    "failure_reasons": [],
+    "verdict": "mtp_depth_wins",
+    "artifact_fingerprint": "sha256:069c2de291fd15b130383119b13f60c45e0f78481a180cf51887ae87c3986b12"
   },
   "verified_on": {
+    "timestamp": "2026-08-19T23:10:32-07:00",
     "hardware": "macOS-26.3.1-arm64-arm-64bit-Mach-O",
     "machine_arch": "arm64",
     "macos": "26.3.1",
-    "model": "Qwen3.8-27B-MTPLX-Optimized-Speed",
-    "timestamp": "2026-08-14T11:28:40-07:00"
+    "model": "Qwen3.8-27B-MTPLX-Optimized-Speed-FP16"
   }
 }