Update model_manager.py

model_manager.py

@@ -51,86 +51,70 @@ class ModelDef:
 CATALOG: list[ModelDef] = [
 
     # ═══════════════════════════════════════════════════════════════════════
-    # TIER 1 — الخيار الافتراضي: جودة عالية + استقرار مُثبَت
+    # TIER 1 — DEFAULT: Qwen3-8B Q4_K_M
+    # Verified repo: unsloth/Qwen3-8B-GGUF (5.03 GB, ~8-12 tok/s on CPU)
+    # Exact filename confirmed on HF Hub: Qwen3-8B-Q4_K_M.gguf
     # ═══════════════════════════════════════════════════════════════════════
     ModelDef(
         id           = "qwen3-8b",
-        name         = "Qwen3 8B — Thinking (bartowski Q4_K_M)",
-        repo         = "bartowski/Qwen3-8B-Instruct-GGUF",
-        glob_primary = "*Q4_K_M*",
-        glob_fallback= "*Q4_K_S*",
-        size_gb      = 5.2,
+        name         = "Qwen3 8B Q4_K_M (unsloth)",
+        repo         = "unsloth/Qwen3-8B-GGUF",
+        glob_primary = "Qwen3-8B-Q4_K_M.gguf",
+        glob_fallback= "Qwen3-8B-Q4_K_S.gguf",
+        size_gb      = 5.1,
         ctx_size     = 16384,
         quality      = 88,
-        has_vision   = True,
+        has_vision   = False,
         has_thinking = True,
-        mmproj_repo  = "bartowski/Qwen3-8B-Instruct-GGUF",
-        mmproj_glob  = "*mmproj*",
+        mmproj_repo  = "",
+        mmproj_glob  = "",
         extra_flags  = "--jinja --reasoning-format deepseek",
         temp=0.6, top_k=20, top_p=0.95, min_p=0.0,
-        notes="الخيار الافتراضي. Qwen3 (مش 3.5) يتجنب KV cache reuse bug. "
-              "bartowski Q4_K_M: KLD=0.0087 — أفضل دقة في هذا الحجم.",
+        notes="Default. Verified filename on HF Hub. 5.03GB. ~8-12 tok/s on CPU.",
     ),
 
     # ═══════════════════════════════════════════════════════════════════════
-    # TIER 2 — أسرع للمهام البسيطة
+    # TIER 2 — FAST: Qwen3-4B Q4_K_M (fits in 4GB, ~20-30 tok/s)
+    # Verified repo: unsloth/Qwen3-4B-GGUF (2.5 GB)
+    # Exact filename confirmed: Qwen3-4B-Q4_K_M.gguf
     # ═══════════════════════════════════════════════════════════════════════
     ModelDef(
         id           = "qwen3-4b",
-        name         = "Qwen3 4B — Fast (bartowski Q4_K_M)",
-        repo         = "bartowski/Qwen3-4B-Instruct-GGUF",
-        glob_primary = "*Q4_K_M*",
-        glob_fallback= "*Q4_K_S*",
-        size_gb      = 2.9,
+        name         = "Qwen3 4B Q4_K_M (unsloth)",
+        repo         = "unsloth/Qwen3-4B-GGUF",
+        glob_primary = "Qwen3-4B-Q4_K_M.gguf",
+        glob_fallback= "Qwen3-4B-Q4_K_S.gguf",
+        size_gb      = 2.5,
         ctx_size     = 16384,
         quality      = 76,
-        has_vision   = True,
+        has_vision   = False,
         has_thinking = True,
-        mmproj_repo  = "bartowski/Qwen3-4B-Instruct-GGUF",
-        mmproj_glob  = "*mmproj*",
+        mmproj_repo  = "",
+        mmproj_glob  = "",
         extra_flags  = "--jinja --reasoning-format deepseek",
         temp=0.6, top_k=20, top_p=0.95, min_p=0.0,
-        notes="أسرع 2x من 8B. للمهام البسيطة والـ classification والـ routing.",
+        notes="Fast fallback. 2.5GB. ~20-30 tok/s on CPU. Good for simple tasks.",
     ),
 
     # ═══════════════════════════════════════════════════════════════════════
-    # TIER 3 — جودة أعلى لما RAM يسمح
+    # TIER 3 — PREMIUM: Qwen3-14B (needs 10GB+ RAM)
     # ═══════════════════════════════════════════════════════════════════════
     ModelDef(
         id           = "qwen3-14b",
-        name         = "Qwen3 14B — Premium (bartowski Q4_K_M)",
-        repo         = "bartowski/Qwen3-14B-Instruct-GGUF",
-        glob_primary = "*Q4_K_M*",
-        glob_fallback= "*Q4_K_S*",
+        name         = "Qwen3 14B Q4_K_M (unsloth)",
+        repo         = "unsloth/Qwen3-14B-GGUF",
+        glob_primary = "Qwen3-14B-Q4_K_M.gguf",
+        glob_fallback= "*Q4_K_M*",
         size_gb      = 9.5,
-        ctx_size     = 12288,   # أصغر context لأن الأوزان أكبر
+        ctx_size     = 12288,
         quality      = 94,
-        has_vision   = True,
-        has_thinking = True,
-        mmproj_repo  = "bartowski/Qwen3-14B-Instruct-GGUF",
-        mmproj_glob  = "*mmproj*",
-        extra_flags  = "--jinja --reasoning-format deepseek",
-        temp=0.6, top_k=20, top_p=0.95, min_p=0.0,
-        notes="Premium tier. يحتاج 16GB صافي (HF CPU Pro). لو RAM متاح اختاره.",
-    ),
-
-    # ═══════════════════════════════════════════════════════════════════════
-    # TIER 4 — المستقبل (تراقبهم model_manager أسبوعياً)
-    # ═══════════════════════════════════════════════════════════════════════
-    ModelDef(
-        id           = "qwen3-32b-q2",
-        name         = "Qwen3 32B — IQ2_XXS (Future — 24GB+)",
-        repo         = "bartowski/Qwen3-32B-Instruct-GGUF",
-        glob_primary = "*IQ2_XXS*",
-        glob_fallback= "*Q2_K*",
-        size_gb      = 12.0,
-        ctx_size     = 8192,
-        quality      = 98,
         has_vision   = False,
         has_thinking = True,
+        mmproj_repo  = "",
+        mmproj_glob  = "",
         extra_flags  = "--jinja --reasoning-format deepseek",
         temp=0.6, top_k=20, top_p=0.95, min_p=0.0,
-        notes="للمستقبل لو ترقيت لـ CPU Pro. يحتاج >14GB RAM صافي.",
+        notes="Premium. Needs 16GB+ RAM. ~4-6 tok/s on CPU.",
     ),
 ]