[
 {
  "id": "qwen3.5-9b",
  "name": "Qwen3.5 9B",
  "maker": "Alibaba Qwen",
  "jp": false,
  "repo": "unsloth/Qwen3.5-9B-GGUF",
  "base": "Qwen/Qwen3.5-9B",
  "file": "Qwen3.5-9B-Q4_K_M.gguf",
  "quant": "Q4_K_M",
  "license": "Apache-2.0",
  "group": "vram",
  "think": "off",
  "args": [
   "--chat-template-kwargs",
   "{\"enable_thinking\":false}"
  ],
  "why": "2026年2月公開の Qwen3.5 系で、12GB に余裕で載る最大級の密モデル。HFのダウンロード数が上位"
 },
 {
  "id": "gemma4-12b",
  "name": "Gemma 4 12B it",
  "maker": "Google",
  "jp": false,
  "repo": "unsloth/gemma-4-12b-it-GGUF",
  "base": "google/gemma-4-12B-it",
  "file": "gemma-4-12b-it-Q4_K_M.gguf",
  "quant": "Q4_K_M",
  "license": "Apache-2.0",
  "group": "vram",
  "think": "off",
  "args": [
   "--chat-template-kwargs",
   "{\"enable_thinking\":false}"
  ],
  "why": "2026年5月公開。多言語に強いとされる Gemma 系の、12GB 向け中核サイズ"
 },
 {
  "id": "gemma4-e4b",
  "name": "Gemma 4 E4B it",
  "maker": "Google",
  "jp": false,
  "repo": "unsloth/gemma-4-E4B-it-GGUF",
  "base": "google/gemma-4-E4B-it",
  "file": "gemma-4-E4B-it-Q4_K_M.gguf",
  "quant": "Q4_K_M",
  "license": "Apache-2.0",
  "group": "vram",
  "think": "off",
  "args": [
   "--chat-template-kwargs",
   "{\"enable_thinking\":false}"
  ],
  "why": "端末向けの小型版。速さの上限側の比較用"
 },
 {
  "id": "ministral3-14b",
  "name": "Ministral 3 14B Instruct 2512",
  "maker": "Mistral AI",
  "jp": false,
  "repo": "mistralai/Ministral-3-14B-Instruct-2512-GGUF",
  "base": "mistralai/Ministral-3-14B-Instruct-2512",
  "file": "Ministral-3-14B-Instruct-2512-Q4_K_M.gguf",
  "quant": "Q4_K_M",
  "license": "Apache-2.0",
  "group": "vram",
  "think": "none",
  "args": [],
  "why": "Mistral 系で 12GB に載る最大の指示モデル。公式 GGUF"
 },
 {
  "id": "llmjp41-8b",
  "name": "llm-jp-4.1 8B thinking",
  "maker": "NII LLM-jp",
  "jp": true,
  "repo": "llm-jp/llm-jp-4.1-8b-thinking-gguf",
  "base": "llm-jp/llm-jp-4.1-8b-thinking",
  "file": "llm-jp-4.1-8b-thinking-Q4_K_M.gguf",
  "quant": "Q4_K_M",
  "license": "Apache-2.0",
  "group": "vram",
  "think": "on",
  "args": [
   "--reasoning-effort",
   "low",
   "--reasoning-budget",
   "1024",
   "--reasoning-budget-message",
   "\n考える時間は終わりです。ここまでの考えで答えます。\n"
  ],
  "why": "国立情報学研究所の日本語フルスクラッチ系の最新（2026年9月）。公式 GGUF。思考モデルなので effort=low",
  "replaced": "llm-jp-4 8B instruct（mmnga-o/llm-jp-4-8b-instruct-gguf Q4_K_M）は llama.cpp b11852 のチャット出力解析（harmony 形式）でエラー 500 になり測定できなかったため、公式 GGUF のある 4.1 thinking に差し替え"
 },
 {
  "id": "qwen3-swallow-8b",
  "name": "Qwen3-Swallow 8B RL v0.2",
  "maker": "Science Tokyo / AIST",
  "jp": true,
  "repo": "mmnga-o/Qwen3-Swallow-8B-RL-v0.2-gguf",
  "base": "tokyotech-llm/Qwen3-Swallow-8B-RL-v0.2",
  "file": "Qwen3-Swallow-8B-RL-v0.2-Q4_K_M.gguf",
  "quant": "Q4_K_M",
  "license": "Apache-2.0",
  "group": "vram",
  "think": "on",
  "args": [
   "--reasoning-budget",
   "1024",
   "--reasoning-budget-message",
   "\n考える時間は終わりです。ここまでの考えで答えます。\n"
  ],
  "why": "Swallow の最新世代（Qwen3 ベースの日本語継続学習）。思考を切れない設計"
 },
 {
  "id": "llama31-swallow-8b",
  "name": "Llama 3.1 Swallow 8B Instruct v0.5",
  "maker": "Science Tokyo / AIST",
  "jp": true,
  "repo": "mmnga/Llama-3.1-Swallow-8B-Instruct-v0.5-gguf",
  "base": "tokyotech-llm/Llama-3.1-Swallow-8B-Instruct-v0.5",
  "file": "Llama-3.1-Swallow-8B-Instruct-v0.5-Q4_K_M.gguf",
  "quant": "Q4_K_M",
  "license": "Llama 3.3 Community License + Gemma Terms of Use",
  "group": "vram",
  "think": "none",
  "args": [],
  "why": "日本語特化 8B の定番。思考なしの Swallow として比較"
 },
 {
  "id": "shisa21-phi4-14b",
  "name": "Shisa V2.1 unphi4 14B",
  "maker": "Shisa.AI",
  "jp": true,
  "repo": "mradermacher/shisa-v2.1-unphi4-14b-GGUF",
  "base": "shisa-ai/shisa-v2.1-unphi4-14b",
  "file": "shisa-v2.1-unphi4-14b.Q4_K_M.gguf",
  "quant": "Q4_K_M",
  "license": "MIT",
  "group": "vram",
  "think": "none",
  "args": [],
  "why": "日英バイリンガル調整の 14B。12GB に載る日本語特化系で最大級"
 },
 {
  "id": "sarashina22-3b",
  "name": "Sarashina2.2 3B instruct v0.1",
  "maker": "SB Intuitions",
  "jp": true,
  "repo": "mmnga/sarashina2.2-3b-instruct-v0.1-gguf",
  "base": "sbintuitions/sarashina2.2-3b-instruct-v0.1",
  "file": "sarashina2.2-3b-instruct-v0.1-Q4_K_M.gguf",
  "quant": "Q4_K_M",
  "license": "MIT",
  "group": "vram",
  "think": "none",
  "args": [],
  "why": "Sarashina 系で重みが公開されている最新の指示モデル（Sarashina3 は重み非公開）"
 },
 {
  "id": "bonsai2-27b",
  "name": "Ternary Bonsai 2 27B",
  "maker": "PrismML",
  "jp": false,
  "repo": "prism-ml/Ternary-Bonsai-2-27B-gguf",
  "base": "Qwen/Qwen3.8-27B (派生)",
  "file": "Ternary-Bonsai-2-27B-PTQ1_0.gguf",
  "quant": "PTQ1_0 (三値)",
  "license": "Apache-2.0",
  "group": "vram",
  "think": "off",
  "args": [
   "--chat-template-kwargs",
   "{\"enable_thinking\":false}"
  ],
  "why": "27B を三値化して 5.95GB にしたという話題のモデル。実在・ライセンスを確認して試す"
 },
 {
  "id": "gpt-oss-20b",
  "name": "gpt-oss-20b",
  "maker": "OpenAI",
  "jp": false,
  "repo": "lmstudio-community/gpt-oss-20b-GGUF",
  "base": "openai/gpt-oss-20b",
  "file": "gpt-oss-20b-MXFP4.gguf",
  "quant": "MXFP4",
  "license": "Apache-2.0",
  "group": "offload",
  "think": "on",
  "args": [
   "--reasoning-effort",
   "low",
   "--reasoning-budget",
   "1024",
   "--reasoning-budget-message",
   "\nTime is up. Answer now.\n"
  ],
  "local": "%USERPROFILE%/.lmstudio/models/lmstudio-community/gpt-oss-20b-GGUF/gpt-oss-20b-MXFP4.gguf",
  "why": "人気の MoE。12.1GB で 12GB に収まりきらず一部が CPU 側に回る"
 },
 {
  "id": "qwen3.6-35b-a3b",
  "name": "Qwen3.6 35B-A3B",
  "maker": "Alibaba Qwen",
  "jp": false,
  "repo": "unsloth/Qwen3.6-35B-A3B-GGUF",
  "base": "Qwen/Qwen3.6-35B-A3B",
  "file": "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf",
  "quant": "UD-Q4_K_M",
  "license": "Apache-2.0",
  "group": "offload",
  "think": "off",
  "args": [
   "--chat-template-kwargs",
   "{\"enable_thinking\":false}"
  ],
  "why": "12GB 機で「専門家を CPU に逃がして動かす」定番の MoE。22GB のうち大半が CPU 側"
 }
]