ltcai 11.1.0 → 11.2.0

package/latticeai/setup/auto_setup.py

@@ -399,36 +399,36 @@ class Recommendation:
         return asdict(self)
 
 
-# 모델 카탈로그. PPT 슬라이드 16 의 "추천 모델" 열과 동기화.
+# 모델 카탈로그 — RAM/VRAM 내림차순. 첫 번째로 조건을 만족하는 항목이 선택된다.
+# 모든 repo id 는 2026-08-10 Hugging Face API 로 확인했다(존재·비gated·정확한 대소문자).
+# OS 오버헤드(~4-6 GB) + KV 캐시 여유를 감안한 보수적 RAM 임계값.
 _MODEL_CATALOG: List[Dict[str, Any]] = [
-    # (min_ram_mb, min_vram_mb, model_id, quant, runtime_preference)
-    # OS 오버헤드(~4-6 GB) + KV 캐시 여유를 감안한 보수적 RAM 임계값
+    # 대형 (48GB+): 최상위 dense / Gemma 4 최대 모델
     {"ram": 64 * 1024, "vram": 32 * 1024,
      "id": "mlx-community/gemma-4-31b-it-4bit", "q": "4bit", "multimodal": True},
-    {"ram": 64 * 1024, "vram": 32 * 1024,
-     "id": "Qwen/Qwen3-VL-30B-A3B-Instruct", "q": "q4_K_M", "multimodal": True},
     {"ram": 48 * 1024, "vram": 24 * 1024,
-     "id": "mlx-community/gemma-4-31b-it-4bit", "q": "4bit", "multimodal": True},
+     "id": "mlx-community/Qwen3.6-27B-4bit", "q": "4bit", "multimodal": True},
+    # MoE (32GB+)
+    {"ram": 40 * 1024, "vram": 24 * 1024,
+     "id": "mlx-community/Qwen3.6-35B-A3B-4bit", "q": "4bit", "multimodal": True},
     {"ram": 32 * 1024, "vram": 16 * 1024,
      "id": "mlx-community/gemma-4-26b-a4b-it-4bit", "q": "4bit", "multimodal": True},
-    {"ram": 48 * 1024, "vram": 24 * 1024,
-     "id": "Qwen/Qwen3-VL-30B-A3B-Instruct", "q": "q4_K_M", "multimodal": True},
-    {"ram": 24 * 1024, "vram": 12 * 1024,
-     "id": "mlx-community/Llama-4-Scout-17B-16E-Instruct-4bit", "q": "4bit", "multimodal": True},
-    {"ram": 16 * 1024, "vram": 8 * 1024,
-     "id": "mlx-community/gemma-4-12b-it-4bit", "q": "4bit", "multimodal": True},
-    {"ram": 32 * 1024, "vram": 16 * 1024,
-     "id": "Qwen/Qwen3-VL-8B-Instruct", "q": "q5_K_M", "multimodal": True},
+    # 범용 (24GB)
     {"ram": 24 * 1024, "vram": 12 * 1024,
-     "id": "Qwen/Qwen3-VL-8B-Instruct", "q": "q4_K_M", "multimodal": True},
+     "id": "mlx-community/gpt-oss-20b-MXFP4-Q8", "q": "MXFP4-Q8", "multimodal": False},
+    # 중형 (16GB)
     {"ram": 16 * 1024, "vram": 8 * 1024,
-     "id": "Qwen/Qwen3-VL-8B-Instruct", "q": "q4_K_M", "multimodal": True},
+     "id": "mlx-community/gemma-4-12B-it-4bit", "q": "4bit", "multimodal": True},
     {"ram": 12 * 1024, "vram": 6 * 1024,
-     "id": "Qwen/Qwen3-VL-4B-Instruct", "q": "q4_K_M", "multimodal": True},
+     "id": "mlx-community/Qwen3.5-9B-MLX-4bit", "q": "4bit", "multimodal": True},
+    # 경량 (10GB) / 초경량 (8GB 이하)
+    {"ram": 10 * 1024, "vram": 6 * 1024,
+     "id": "mlx-community/gemma-4-e4b-it-4bit", "q": "4bit", "multimodal": True},
     {"ram":  8 * 1024, "vram": 4 * 1024,
-     "id": "Qwen/Qwen3-VL-4B-Instruct", "q": "q4_K_M", "multimodal": True},
+     "id": "mlx-community/gemma-4-e2b-it-4bit", "q": "4bit", "multimodal": True},
+    # 최소 사양: 사진은 못 읽지만 한국어로 대답은 한다.
     {"ram":  4 * 1024, "vram": 0,
-     "id": "Qwen/Qwen3-VL-4B-Instruct", "q": "q4_K_M", "multimodal": True},
+     "id": "mlx-community/LFM2.5-2.6B-4bit", "q": "4bit", "multimodal": False},
 ]
 
 
@@ -622,17 +622,14 @@ def plan(profile: SystemProfile, rec: Recommendation) -> InstallPlan:
     # 모델 가중치 풀
     model_command = ["huggingface-cli", "download", rec.model_id, "--quiet"]
     if rec.runtime == "ollama":
-        lower = rec.model_id.lower()
-        if "gemma-4-31b" in lower:
-            model_command = ["ollama", "pull", "hf.co/ggml-org/gemma-4-31B-it-GGUF:Q4_K_M"]
-        elif "gemma-4-12b" in lower:
-            model_command = ["ollama", "pull", "hf.co/ggml-org/gemma-4-12B-it-GGUF:Q4_K_M"]
-        elif "llama-4-scout" in lower:
-            model_command = ["ollama", "pull", "hf.co/ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF:Q4_K_M"]
-        elif "qwen3-vl-8b" in lower:
-            model_command = ["ollama", "pull", "qwen3-vl:8b"]
-        elif "qwen3-vl-4b" in lower:
-            model_command = ["ollama", "pull", "qwen3-vl:4b"]
+        # Local import: this module keeps a stdlib-only import graph so it can be
+        # run before the app is wired. The alias table is the single source of
+        # truth for per-engine repo ids — the old hardcoded if/elif chain here
+        # drifted from it and still named Llama 4 Scout and Qwen3-VL.
+        from latticeai.services.model_catalog import MODEL_ENGINE_ALIASES
+        ollama_target = (MODEL_ENGINE_ALIASES.get(rec.model_id.lower()) or {}).get("ollama")
+        if ollama_target:
+            model_command = ["ollama", "pull", ollama_target]
     elif rec.runtime == "lmstudio":
         model_command = ["lms", "get", rec.model_id]
     steps.append(InstallStep(

package/latticeai/setup/wizard.py

@@ -562,47 +562,63 @@ def scan_environment() -> Dict[str, Any]:
 
 # ── Model Catalog ─────────────────────────────────────────────────────────────
 # (model_id, display_name, size_gb, tag, description, min_ram_gb)
+# 11.2.0: 모든 repo id 를 2026-08-10 Hugging Face API 로 확인했다 — 존재 여부,
+# gated 여부, 정확한 대소문자, siblings 합계 크기. 크기 값은 측정값이다.
+# 삭제: Qwen3-VL 전 라인업(2025-10, Qwen3.5/3.6 이 상위 세대), Llama 4 Scout
+# (vLLM/LM Studio 경로가 gated 인 meta-llama 저장소를 가리켰다).
 _MODEL_CATALOG = [
-    ("mlx-community/Qwen3-VL-4B-Instruct-4bit",      "Qwen3-VL 4B",          2.7,  "VLM",     "최신 Qwen 멀티모달 · 저사양",       4),
-    ("mlx-community/gemma-4-e2b-it-4bit",            "Gemma 4 E2B",          3.6,  "VLM",     "Gemma 4 소형 멀티모달",            8),
-    ("mlx-community/Qwen3-VL-8B-Instruct-4bit",      "Qwen3-VL 8B",          4.8,  "VLM",     "최신 Qwen 멀티모달 · 균형 추천",    16),
-    ("mlx-community/gemma-4-12b-it-4bit",            "Gemma 4 12B",          7.6,  "VLM",     "Gemma 4 기본 추천 · 4bit",         16),
-    ("mlx-community/Llama-4-Scout-17B-16E-Instruct-4bit", "Llama 4 Scout",   11.8, "VLM",     "Meta 최신 멀티모달 Scout",          24),
-    ("mlx-community/gemma-4-26b-a4b-it-4bit",        "Gemma 4 26B",         15.6,  "VLM",     "이미지 지원 · 대형 추천",        32),
-    ("mlx-community/gemma-4-31b-it-4bit",            "Gemma 4 31B",         18.4,  "VLM+",    "Gemma 4 최신 31B instruct",      48),
-    ("mlx-community/Qwen3-VL-30B-A3B-Instruct-4bit", "Qwen3-VL 30B A3B",    18.0,  "VLM+",    "최신 Qwen 대형 멀티모달",        48),
+    ("mlx-community/LFM2.5-2.6B-4bit",               "LFM2.5 2.6B",          1.5,  "LLM",     "가장 가벼움 · 한국어 대화 (사진은 못 읽음)", 4),
+    ("mlx-community/gemma-4-e2b-it-4bit",            "Gemma 4 E2B",          3.6,  "VLM",     "사진을 읽는 가장 작은 모델",        8),
+    ("mlx-community/gemma-4-e4b-it-4bit",            "Gemma 4 E4B",          5.2,  "VLM",     "E2B 보다 한 단계 위 · 여전히 가벼움", 10),
+    ("mlx-community/Qwen3.5-9B-MLX-4bit",            "Qwen3.5 9B",           6.0,  "VLM",     "중형 멀티모달 · 균형 추천",         12),
+    ("mlx-community/gemma-4-12B-it-4bit",            "Gemma 4 12B",          6.8,  "VLM",     "Gemma 4 기본 추천 · 4bit",          16),
+    ("mlx-community/gpt-oss-20b-MXFP4-Q8",           "GPT-OSS 20B",         12.1,  "LLM",     "범용 · 다운로드 최다 (사진은 못 읽음)", 24),
+    ("mlx-community/gemma-4-26b-a4b-it-4bit",        "Gemma 4 26B A4B",     15.4,  "VLM",     "MoE · 대형 추천",                   32),
+    ("mlx-community/Qwen3.6-27B-4bit",               "Qwen3.6 27B",         16.1,  "VLM+",    "dense 최상위 · 느리지만 일정함",     48),
+    ("mlx-community/gemma-4-31b-it-4bit",            "Gemma 4 31B",         18.4,  "VLM+",    "Gemma 4 최대 모델",                 48),
+    ("mlx-community/Qwen3.6-35B-A3B-4bit",           "Qwen3.6 35B A3B",     20.4,  "VLM+",    "MoE 최상위",                        48),
 ]
 
 _CROSS_PLATFORM_MODEL_CATALOG: Dict[str, List[Tuple[str, str, float, str, str, int]]] = {
     "ollama": [
-        ("ollama:qwen3-vl:4b", "Qwen3-VL 4B", 2.7, "VLM", "Ollama 멀티모달 · 저사양", 4),
-        ("ollama:qwen3-vl:8b", "Qwen3-VL 8B", 4.8, "VLM", "Ollama 멀티모달 · 균형 추천", 16),
+        ("ollama:hf.co/LiquidAI/LFM2.5-2.6B-GGUF:Q4_K_M", "LFM2.5 2.6B Q4", 1.7, "LLM", "가장 가벼움 · 한국어 대화", 4),
+        ("ollama:hf.co/ggml-org/gemma-4-E2B-it-GGUF:Q4_K_M", "Gemma 4 E2B Q4", 3.8, "VLM", "Hugging Face GGUF 기반 Gemma 4", 8),
+        ("ollama:hf.co/ggml-org/gemma-4-E4B-it-GGUF:Q4_K_M", "Gemma 4 E4B Q4", 5.4, "VLM", "Hugging Face GGUF 기반 Gemma 4", 10),
         ("ollama:hf.co/ggml-org/gemma-4-12B-it-GGUF:Q4_K_M", "Gemma 4 12B Q4", 7.9, "VLM", "Hugging Face GGUF 기반 Gemma 4", 16),
-        ("ollama:hf.co/ggml-org/gemma-4-31B-it-GGUF:Q4_K_M", "Gemma 4 31B Q4", 18.7, "VLM+", "Hugging Face GGUF 기반 Gemma 4", 48),
-        ("ollama:hf.co/ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF:Q4_K_M", "Llama 4 Scout Q4", 12.0, "VLM", "Meta 최신 멀티모달 Scout", 24),
+        ("ollama:hf.co/ggml-org/gpt-oss-20b-GGUF:Q4_K_M", "GPT-OSS 20B Q4", 12.5, "LLM", "범용 · 다운로드 최다", 24),
+        ("ollama:hf.co/ggml-org/gemma-4-26B-A4B-it-GGUF:Q4_K_M", "Gemma 4 26B Q4", 16.0, "VLM", "MoE · 대형 추천", 32),
+        ("ollama:hf.co/ggml-org/Qwen3.6-27B-GGUF:Q4_K_M", "Qwen3.6 27B Q4", 16.6, "VLM+", "dense 최상위", 48),
+        ("ollama:hf.co/ggml-org/gemma-4-31B-it-GGUF:Q4_K_M", "Gemma 4 31B Q4", 18.7, "VLM+", "Gemma 4 최대 모델", 48),
     ],
     "lmstudio": [
-        ("lmstudio:Qwen/Qwen3-VL-4B-Instruct", "Qwen3-VL 4B", 2.7, "VLM", "LM Studio 멀티모달 · 저사양", 4),
-        ("lmstudio:Qwen/Qwen3-VL-8B-Instruct", "Qwen3-VL 8B", 4.8, "VLM", "LM Studio 멀티모달 · 균형 추천", 16),
-        ("lmstudio:ggml-org/gemma-4-12B-it-GGUF", "Gemma 4 12B 4-bit", 7.9, "VLM", "LM Studio GGUF Gemma 4", 16),
-        ("lmstudio:ggml-org/gemma-4-31B-it-GGUF", "Gemma 4 31B 4-bit", 18.7, "VLM+", "LM Studio GGUF Gemma 4", 48),
-        ("lmstudio:Qwen/Qwen3-VL-30B-A3B-Instruct", "Qwen3-VL 30B A3B", 18.0, "VLM+", "대형 Qwen 멀티모달 · 24GB+ VRAM 권장", 32),
-        ("lmstudio:meta-llama/Llama-4-Scout-17B-16E-Instruct", "Llama 4 Scout", 12.0, "VLM", "Meta 최신 멀티모달 Scout", 24),
+        ("lmstudio:LiquidAI/LFM2.5-2.6B-GGUF", "LFM2.5 2.6B", 1.7, "LLM", "가장 가벼움 · 한국어 대화", 4),
+        ("lmstudio:ggml-org/gemma-4-E2B-it-GGUF", "Gemma 4 E2B", 3.8, "VLM", "LM Studio GGUF Gemma 4", 8),
+        ("lmstudio:ggml-org/gemma-4-E4B-it-GGUF", "Gemma 4 E4B", 5.4, "VLM", "LM Studio GGUF Gemma 4", 10),
+        ("lmstudio:ggml-org/gemma-4-12B-it-GGUF", "Gemma 4 12B", 7.9, "VLM", "LM Studio GGUF Gemma 4", 16),
+        ("lmstudio:ggml-org/gpt-oss-20b-GGUF", "GPT-OSS 20B", 12.5, "LLM", "범용 · 다운로드 최다", 24),
+        ("lmstudio:ggml-org/gemma-4-26B-A4B-it-GGUF", "Gemma 4 26B", 16.0, "VLM", "MoE · 대형 추천", 32),
+        ("lmstudio:ggml-org/Qwen3.6-27B-GGUF", "Qwen3.6 27B", 16.6, "VLM+", "dense 최상위", 48),
+        ("lmstudio:ggml-org/gemma-4-31B-it-GGUF", "Gemma 4 31B", 18.7, "VLM+", "Gemma 4 최대 모델", 48),
     ],
     "vllm": [
-        ("vllm:Qwen/Qwen3-VL-4B-Instruct", "Qwen3-VL 4B", 2.7, "VLM", "내 컴퓨터 GPU 실행 도구 권장", 4),
-        ("vllm:Qwen/Qwen3-VL-8B-Instruct", "Qwen3-VL 8B", 4.8, "VLM", "내 컴퓨터 NVIDIA 실행 도구 권장", 16),
-        ("vllm:google/gemma-4-12b-it", "Gemma 4 12B", 7.6, "VLM", "Gemma 4 기본 추천 · 4bit", 16),
-        ("vllm:Qwen/Qwen3-VL-30B-A3B-Instruct", "Qwen3-VL 30B A3B", 18.0, "VLM+", "대형 Qwen 멀티모달 · 24GB+ VRAM 권장", 32),
-        ("vllm:suitch/gemma-4-31B-it-4bit", "Gemma 4 31B", 18.7, "VLM+", "Gemma 4 최신 31B instruct", 48),
-        ("vllm:meta-llama/Llama-4-Scout-17B-16E-Instruct", "Llama 4 Scout", 12.0, "VLM", "Meta 최신 멀티모달 Scout", 24),
+        ("vllm:LiquidAI/LFM2.5-2.6B", "LFM2.5 2.6B", 5.2, "LLM", "가장 가벼움 · 한국어 대화", 8),
+        ("vllm:google/gemma-4-E2B-it", "Gemma 4 E2B", 6.0, "VLM", "내 컴퓨터 GPU 실행 도구 권장", 12),
+        ("vllm:google/gemma-4-E4B-it", "Gemma 4 E4B", 9.0, "VLM", "내 컴퓨터 GPU 실행 도구 권장", 16),
+        ("vllm:Qwen/Qwen3.5-9B", "Qwen3.5 9B", 18.0, "VLM", "중형 멀티모달 · 균형 추천", 24),
+        ("vllm:google/gemma-4-12B-it", "Gemma 4 12B", 24.0, "VLM", "Gemma 4 기본 추천", 32),
+        ("vllm:openai/gpt-oss-20b", "GPT-OSS 20B", 13.5, "LLM", "범용 · 다운로드 최다", 24),
+        ("vllm:Qwen/Qwen3.6-27B", "Qwen3.6 27B", 54.0, "VLM+", "dense 최상위 · 24GB+ VRAM 권장", 64),
+        ("vllm:Qwen/Qwen3.6-35B-A3B", "Qwen3.6 35B A3B", 70.0, "VLM+", "MoE 최상위 · 24GB+ VRAM 권장", 80),
     ],
     "llamacpp": [
-        ("llamacpp:Qwen/Qwen3-VL-4B-Instruct-GGUF", "Qwen3-VL 4B GGUF", 2.7, "GGUF", "CPU/Vulkan 백업 · 멀티모달 GGUF", 4),
-        ("llamacpp:Qwen/Qwen3-VL-8B-Instruct-GGUF", "Qwen3-VL 8B GGUF", 4.8, "GGUF", "CPU/Vulkan 백업 · 균형형", 16),
+        ("llamacpp:LiquidAI/LFM2.5-2.6B-GGUF", "LFM2.5 2.6B GGUF", 1.7, "GGUF", "CPU/Vulkan 백업 · 가장 가벼움", 4),
+        ("llamacpp:ggml-org/gemma-4-E2B-it-GGUF", "Gemma 4 E2B GGUF", 3.8, "GGUF", "Gemma 4 E2B Q4_K_M", 8),
+        ("llamacpp:ggml-org/gemma-4-E4B-it-GGUF", "Gemma 4 E4B GGUF", 5.4, "GGUF", "Gemma 4 E4B Q4_K_M", 10),
         ("llamacpp:ggml-org/gemma-4-12B-it-GGUF", "Gemma 4 12B GGUF", 7.9, "GGUF", "Gemma 4 12B Q4_K_M", 16),
+        ("llamacpp:ggml-org/gpt-oss-20b-GGUF", "GPT-OSS 20B GGUF", 12.5, "GGUF", "범용 · 다운로드 최다", 24),
+        ("llamacpp:ggml-org/gemma-4-26B-A4B-it-GGUF", "Gemma 4 26B GGUF", 16.0, "GGUF", "MoE · 대형 추천", 32),
+        ("llamacpp:ggml-org/Qwen3.6-27B-GGUF", "Qwen3.6 27B GGUF", 16.6, "GGUF", "dense 최상위", 48),
         ("llamacpp:ggml-org/gemma-4-31B-it-GGUF", "Gemma 4 31B GGUF", 18.7, "GGUF", "Gemma 4 31B Q4_K_M", 48),
-        ("llamacpp:ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF", "Llama 4 Scout GGUF", 12.0, "GGUF", "Meta 최신 멀티모달 Scout", 24),
     ],
 }
 
@@ -612,37 +628,51 @@ _VERSIONED_MODEL_PATTERNS = (
     ("llama", re.compile(r"\bllama[-\s]?(\d+(?:\.\d+)?)", re.IGNORECASE)),
 )
 
+# RAM(GB) 내림차순 티어. 첫 번째로 조건을 만족하는 항목이 "이 컴퓨터에서 가장
+# 좋은 모델"이 된다. 초경량(≤8) / 경량(16) / 중형(24) / MoE(32) / 대형(48).
 _BEST_MODEL_TIERS: Dict[str, List[Tuple[int, str]]] = {
     "local_mlx": [
         (48, "mlx-community/gemma-4-31b-it-4bit"),
         (32, "mlx-community/gemma-4-26b-a4b-it-4bit"),
-        (16, "mlx-community/gemma-4-12b-it-4bit"),
-        (16, "mlx-community/Qwen3-VL-8B-Instruct-4bit"),
-        (4, "mlx-community/Qwen3-VL-4B-Instruct-4bit"),
+        (24, "mlx-community/gemma-4-12B-it-4bit"),
+        (16, "mlx-community/Qwen3.5-9B-MLX-4bit"),
+        (8, "mlx-community/gemma-4-e2b-it-4bit"),
+        (4, "mlx-community/LFM2.5-2.6B-4bit"),
     ],
     "ollama": [
         (48, "ollama:hf.co/ggml-org/gemma-4-31B-it-GGUF:Q4_K_M"),
-        (16, "ollama:hf.co/ggml-org/gemma-4-12B-it-GGUF:Q4_K_M"),
-        (16, "ollama:qwen3-vl:8b"),
-        (4, "ollama:qwen3-vl:4b"),
+        (32, "ollama:hf.co/ggml-org/gemma-4-26B-A4B-it-GGUF:Q4_K_M"),
+        (24, "ollama:hf.co/ggml-org/gemma-4-12B-it-GGUF:Q4_K_M"),
+        (16, "ollama:hf.co/ggml-org/gemma-4-E4B-it-GGUF:Q4_K_M"),
+        (8, "ollama:hf.co/ggml-org/gemma-4-E2B-it-GGUF:Q4_K_M"),
+        (4, "ollama:hf.co/LiquidAI/LFM2.5-2.6B-GGUF:Q4_K_M"),
     ],
     "lmstudio": [
         (48, "lmstudio:ggml-org/gemma-4-31B-it-GGUF"),
-        (16, "lmstudio:ggml-org/gemma-4-12B-it-GGUF"),
-        (16, "lmstudio:Qwen/Qwen3-VL-8B-Instruct"),
-        (4, "lmstudio:Qwen/Qwen3-VL-4B-Instruct"),
+        (32, "lmstudio:ggml-org/gemma-4-26B-A4B-it-GGUF"),
+        (24, "lmstudio:ggml-org/gemma-4-12B-it-GGUF"),
+        (16, "lmstudio:ggml-org/gemma-4-E4B-it-GGUF"),
+        (8, "lmstudio:ggml-org/gemma-4-E2B-it-GGUF"),
+        (4, "lmstudio:LiquidAI/LFM2.5-2.6B-GGUF"),
     ],
+    # vLLM serves the upstream bf16 repos, which are 3-4x the MLX 4-bit builds
+    # (Qwen3.6-27B is 54GB, not 16GB), so its thresholds sit far higher than the
+    # local_mlx ones. Picking by RAM alone would nominate a model no consumer
+    # GPU can hold.
     "vllm": [
-        (48, "vllm:suitch/gemma-4-31B-it-4bit"),
-        (16, "vllm:google/gemma-4-12b-it"),
-        (16, "vllm:Qwen/Qwen3-VL-8B-Instruct"),
-        (4, "vllm:Qwen/Qwen3-VL-4B-Instruct"),
+        (96, "vllm:Qwen/Qwen3.6-27B"),
+        (32, "vllm:google/gemma-4-12B-it"),
+        (24, "vllm:Qwen/Qwen3.5-9B"),
+        (16, "vllm:google/gemma-4-E4B-it"),
+        (8, "vllm:LiquidAI/LFM2.5-2.6B"),
     ],
     "llamacpp": [
         (48, "llamacpp:ggml-org/gemma-4-31B-it-GGUF"),
-        (16, "llamacpp:ggml-org/gemma-4-12B-it-GGUF"),
-        (16, "llamacpp:Qwen/Qwen3-VL-8B-Instruct-GGUF"),
-        (4, "llamacpp:Qwen/Qwen3-VL-4B-Instruct-GGUF"),
+        (32, "llamacpp:ggml-org/gemma-4-26B-A4B-it-GGUF"),
+        (24, "llamacpp:ggml-org/gemma-4-12B-it-GGUF"),
+        (16, "llamacpp:ggml-org/gemma-4-E4B-it-GGUF"),
+        (8, "llamacpp:ggml-org/gemma-4-E2B-it-GGUF"),
+        (4, "llamacpp:LiquidAI/LFM2.5-2.6B-GGUF"),
     ],
 }
 
@@ -656,7 +686,10 @@ def _catalog_row_family_version(row: Tuple[str, str, float, str, str, int]) -> T
     for family, pattern in _VERSIONED_MODEL_PATTERNS:
         match = pattern.search(text)
         if match:
-            version = _version_tuple(match.group(1))
+            # Major version only — see model_catalog._model_family_version. The
+            # filter hides superseded *generations*; Qwen3.5 and Qwen3.6 are one
+            # generation filling two different RAM tiers and must coexist.
+            version = _version_tuple(match.group(1))[:1]
             if version:
                 return family, version
     return None

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ltcai",
-  "version": "11.1.0",
+  "version": "11.2.0",
   "description": "Lattice AI — local-first Digital Brain that keeps your knowledge durable across any AI model.",
   "homepage": "https://github.com/TaeSooPark-PTS/LatticeAI#readme",
   "repository": {

package/scripts/check_current_release_docs.mjs

@@ -6,7 +6,7 @@ const root = process.cwd();
 const pkg = JSON.parse(readFileSync(path.join(root, "package.json"), "utf8"));
 const version = pkg.version;
 const releaseDir = `output/release/v${version}`;
-const releaseTheme = "Product Intelligence";
+const releaseTheme = "All Systems On";
 const title = `${version} — ${releaseTheme}`;
 const escapedVersion = version.replaceAll(".", "\\.");
 

package/scripts/check_server_i18n.mjs

@@ -34,6 +34,7 @@ const LOCALIZED = [
   "chat",
   "chat_history",
   "chat_intents",
+  "features",
   "knowledge_graph",
   "local_files",
   "mcp",

package/scripts/release_screen_claims.json

@@ -21,6 +21,19 @@
     "specific, not less."
   ],
 
+  "11.2.0": {
+    "04-brain-chat-home.png": {
+      "_why": [
+        "The dock rail gained a fourth item (기능) opening the feature-toggle",
+        "drawer. The rail change itself is small; the LivingBrain animation",
+        "noise threshold (8%) may exceed it, so the screen is claimed with the",
+        "rail region only."
+      ],
+      "region": [0.0, 0.25, 0.12, 0.85],
+      "min_pct": 0.5
+    }
+  },
+
   "11.1.0": {
     "_why": [
       "Feature release with no intentional change to the twelve captured",