更新于 2026-10-09
模型排行榜的 本地 LLM
所有可在本地运行的 LLM，按照针对 34 个模型进行基准测试的透明综合评分排序（总计 253 个）。
结论： Phi-4 14B 以 69.1/100 的综合评分位居该排行榜榜首。要检查您的 GPU 能容纳哪些模型，请使用 显存计算器.
您是在比较两个具体模型吗？请使用 对比工具. 您是根据硬件还是使用场景来选择？参见 排名。云端与本地的成本如何比较？请使用 成本计算器.
| # | 模型 | 评分 | Params | 最小 VRAM（Q4） | 许可证 | 上下文 | 关键基准测试 |
|---|---|---|---|---|---|---|---|
| 1 | Phi-4 14BPhi | 69,1 | 14B | 9 GB | MIT | 16K | MMLU 85 |
| 2 | Phi-3.5 MiniPhi | 67,2 | 3.8B | 3 GB | MIT | 131K | MMLU 69 |
| 3 | Qwen 2.5 7BQwen | 67,0 | 7B | 5 GB | Apache 2.0 | 131K | MMLU 74 |
| 4 | Granite 4.1 8B InstructGranite | 66,5 | 8B | 5 GB | Apache 2.0 | 131K | MMLU 74 |
| 5 | Qwen 2.5 3B InstructQwen | 65,8 | 3B | 2 GB | Qwen Research License | 32K | MMLU 66 |
| 6 | Qwen 2.5 14B InstructQwen | 65,7 | 14B | 9 GB | Apache 2.0 | 131K | MMLU 80 |
| 7 | Phi-4 Mini 3.8BPhi | 64,8 | 3.8B | 3 GB | MIT | 128K | MMLU 67 |
| 8 | DeepSeek R1 671BDeepSeek | 63,0 | 671B | 400 GB | MIT | 128K | MMLU 91 |
| 9 | SmolLM3 3BSmolLM | 61,6 | 3B | 2 GB | Apache 2.0 | 128K | MMLU 60 |
| 10 | Mistral Small 3.1 24BMistral | 61,4 | 24B | 14 GB | Apache 2.0 | 128K | MMLU 81 |
| 11 | Llama 3.1 8BLlama | 60,5 | 8B | 6 GB | Llama 3 Community | 131K | MMLU 73 |
| 12 | Qwen 2.5 32BQwen | 60,5 | 32B | 19 GB | Apache 2.0 | 131K | MMLU 83 |
| 13 | Mistral Small 3Mistral | 59,9 | 24B | 14 GB | Apache 2.0 | 32K | MMLU 81 |
| 14 | Llama 3.2 3BLlama | 59,5 | 3B | 3 GB | Llama 3 Community | 131K | MMLU 63 |
| 15 | Granite 3.2 8B InstructGranite | 57,5 | 8B | 5 GB | Apache 2.0 | 128K | MMLU 67 |
| 16 | Llama 3.1 405B InstructLlama | 57,5 | 405B | 240 GB | Llama 3.1 Community | 128K | MMLU 89 |
| 17 | Llama 3.1 70BLlama | 55,9 | 70B | 40 GB | Llama 3 Community | 131K | MMLU 86 |
| 18 | Llama 3.3 70B InstructLlama | 55,9 | 70B | 40 GB | Llama 3.3 Community | 128K | MMLU 86 |
| 19 | Qwen 2.5 72B InstructQwen | 55,9 | 72B | 42 GB | Qwen 许可证 | 131K | MMLU 86 |
| 20 | Mistral Nemo 12B InstructMistral | 55,8 | 12B | 7 GB | Apache 2.0 | 128K | MMLU 68 |
| 21 | Gemma 3 27BGemma | 55,4 | 27B | 16 GB | Gemma | 128K | MMLU 79 |
| 22 | Falcon 3 10B InstructFalcon | 54,6 | 10B | 6 GB | TII Falcon-LLM 许可证 2.0 | 32K | MMLU 73 |
| 23 | Gemma 2 9BGemma | 54,3 | 9B | 6 GB | Gemma | 8K | MMLU 71 |
| 24 | Falcon 3 7B InstructFalcon | 53,1 | 7B | 5 GB | TII Falcon-LLM 许可证 2.0 | 32K | MMLU 71 |
| 25 | Yi 1.5 34B ChatYi | 48,0 | 34B | 20 GB | Apache 2.0 | 4K | MMLU 77 |
| 26 | Mixtral 8x22B InstructMistral | 47,4 | 141B | 82 GB | Apache 2.0 | 64K | MMLU 78 |
| 27 | Gemma 2 27BGemma | 47,3 | 27B | 16 GB | Gemma | 8K | MMLU 75 |
| 28 | Mistral 7B InstructMistral | 46,4 | 7B | 5 GB | Apache 2.0 | 32K | MMLU 60 |
| 29 | Gemma 2 2BGemma | 45,3 | 2B | 2 GB | Gemma | 8K | MMLU 52 |
| 30 | Falcon Mamba 7BFalcon | 39,9 | 7B | 5 GB | TII Falcon-LLM 许可证 2.0 | 8K | MMLU 62 |
| 31 | Mixtral 8x7BMistral | 39,5 | 47B | 26 GB | Apache 2.0 | 32K | MMLU 71 |
| 32 | OLMoE 1B-7B InstructOLMo | 34,8 | 7B | 4 GB | Apache 2.0 | 4K | MMLU 52 |
| 33 | DBRX InstructDBRX | 34,7 | 132B | 76 GB | Databricks Open Model License | 32K | MMLU 74 |
| 34 | Snowflake Arctic InstructArctic | 30,7 | 480B | 290 GB | Apache 2.0 | 4K | MMLU 67 |
| — | Lucie 7BLucie | n/a | 7B | 5 GB | Apache 2.0 | 4K | — |
| — | CroissantLLM 1.3BCroissant | n/a | 1.3B | 1 GB | MIT | 2K | — |
| — | Qwen 2.5 Coder 7BQwen | n/a | 7B | 5 GB | Apache 2.0 | 131K | — |
| — | Qwen 2.5 Coder 32BQwen | n/a | 32B | 19 GB | Apache 2.0 | 131K | — |
| — | DeepSeek R1 Distill 7BDeepSeek | n/a | 7B | 5 GB | MIT | 32K | — |
| — | DeepSeek R1 Distill 32BDeepSeek | n/a | 32B | 19 GB | MIT | 32K | — |
| — | DeepSeek Coder V2 Lite 16BDeepSeek | n/a | 16B | 10 GB | MIT | 131K | — |
| — | Llama 3.2 Vision 11BLlama | n/a | 11B | 8 GB | Llama 3 Community | 131K | — |
| — | Qwen 2 VL 7BQwen | n/a | 7B | 6 GB | Apache 2.0 | 32K | — |
| — | Qwen 3 8BQwen | n/a | 8B | 5 GB | Apache 2.0 | 131K | — |
| — | Qwen 3 14BQwen | n/a | 14B | 9 GB | Apache 2.0 | 131K | — |
| — | Qwen 3 32BQwen | n/a | 32B | 19 GB | Apache 2.0 | 131K | — |
| — | Qwen 3 235B-A22BQwen | n/a | 235B | 142 GB | Apache 2.0 | 131K | — |
| — | Qwen 2.5 VL 7BQwen | n/a | 7B | 6 GB | Apache 2.0 | 128K | — |
| — | Qwen 2.5 VL 72BQwen | n/a | 72B | 42 GB | Qwen 许可证 | 128K | — |
| — | Qwen 2.5 Omni 7BQwen | n/a | 7B | 6 GB | Apache 2.0 | 32K | — |
| — | QwQ 32BQwen | n/a | 32B | 19 GB | Apache 2.0 | 131K | — |
| — | DeepSeek R1 Distill Llama 70BDeepSeek | n/a | 70B | 40 GB | Llama 3.3 Community + DeepSeek | 128K | — |
| — | DeepSeek V3 671BDeepSeek | n/a | 671B | 400 GB | DeepSeek License | 128K | — |
| — | Gemma 3 4BGemma | n/a | 4B | 4 GB | Gemma | 128K | — |
| — | Gemma 3 12BGemma | n/a | 12B | 7 GB | Gemma | 128K | — |
| — | Phi-4 Multimodal 5.6BPhi | n/a | 5.6B | 4 GB | MIT | 128K | — |
| — | Phi-4 Reasoning 14BPhi | n/a | 14B | 9 GB | MIT | 32K | — |
| — | Command R+ 104B (08-2024)Command | n/a | 104B | 60 GB | CC-BY-NC 4.0 | 128K | — |
| — | Aya Expanse 8BAya | n/a | 8B | 5 GB | CC-BY-NC 4.0 | 8K | — |
| — | Aya Expanse 32BAya | n/a | 32B | 19 GB | CC-BY-NC 4.0 | 8K | — |
| — | EuroLLM 9B InstructEuroLLM | n/a | 9B | 6 GB | Apache 2.0 | 4K | — |
| — | Teuken 7B InstructTeuken | n/a | 7B | 5 GB | Apache 2.0（商业用途） | 4K | — |
| — | Pleias 3B PreviewPleias | n/a | 3B | 2 GB | Apache 2.0 | 2K | — |
| — | Pleias-RAG 1BPleias | n/a | 1.2B | 1 GB | Apache 2.0 | 2K | — |
| — | Moshi 7BMoshi | n/a | 7.6B | 5 GB | CC-BY 4.0 | 4K | — |
| — | Helium 1 2BHelium | n/a | 2B | 2 GB | CC-BY-SA 4.0 | 4K | — |
| — | SmolLM2 1.7B InstructSmolLM | n/a | 1.7B | 1 GB | Apache 2.0 | 8K | — |
| — | SmolVLM2 2.2B InstructSmolLM | n/a | 2.2B | 2 GB | Apache 2.0 | 8K | — |
| — | GLM-5.1GLM | n/a | 744B | 445 GB | MIT | 200K | — |
| — | MiniMax-M2.7MiniMax | n/a | 229B | 138 GB | Apache 2.0 | 205K | — |
| — | Gemma 4 31BGemma | n/a | 31B | 18 GB | Apache 2.0 | 256K | — |
| — | Gemma 4 E4BGemma | n/a | 4B | 10 GB | Apache 2.0 | 128K | — |
| — | Qwen 3.5 9BQwen | n/a | 9B | 6 GB | Apache 2.0 | 262K | — |
| — | Qwen 3.5 27BQwen | n/a | 27B | 16 GB | Apache 2.0 | 262K | — |
| — | Qwen 3.5 397B-A17BQwen | n/a | 397B | 240 GB | Apache 2.0 | 262K | — |
| — | Qwen 3.6 35B-A3BQwen | n/a | 35B | 21 GB | Apache 2.0 | 262K | — |
| — | Qwen3-Coder-Next 80B-A3BQwen | n/a | 80B | 48 GB | Apache 2.0 | 262K | — |
| — | Qwen3-Coder 30B-A3BQwen | n/a | 30B | 19 GB | Apache 2.0 | 262K | — |
| — | Mistral Small 4Mistral | n/a | 119B | 72 GB | Apache 2.0 | 256K | — |
| — | Devstral Small 2 24BMistral | n/a | 24B | 14 GB | Apache 2.0 | 256K | — |
| — | Voxtral-4B-TTSMistral | n/a | 4B | 3 GB | CC-BY-NC 4.0 | 4K | — |
| — | DeepSeek R2 32BDeepSeek | n/a | 32B | 19 GB | MIT | 128K | — |
| — | DeepSeek V3.2DeepSeek | n/a | 685B | 410 GB | MIT | 128K | — |
| — | Kimi K2.5Kimi | n/a | 1000B | 600 GB | 修改版 MIT | 256K | — |
| — | Nemotron 3 Super 120BNemotron | n/a | 120B | 72 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | OLMo 3 7BOLMo | n/a | 7B | 5 GB | Apache 2.0 | 8K | — |
| — | OLMo 3 32BOLMo | n/a | 32B | 19 GB | Apache 2.0 | 65K | — |
| — | Tiny Aya 3.35BAya | n/a | 3.35B | 2 GB | CC-BY-NC 4.0 | 8K | — |
| — | Granite 4.0 3B VisionGranite | n/a | 3B | 2 GB | Apache 2.0 | 16K | — |
| — | Step 3.5 FlashStep | n/a | 196B | 118 GB | Apache 2.0 | 256K | — |
| — | Falcon H1R 7BFalcon | n/a | 7B | 5 GB | TII Falcon-LLM 许可证 2.0 | 32K | — |
| — | Mistral Small 3.2 24BMistral | n/a | 24B | 14 GB | Apache 2.0 | 128K | — |
| — | Codestral 22B v0.1Mistral | n/a | 22B | 13 GB | Mistral Non-Production License | 32K | — |
| — | Codestral Mamba 7BMistral | n/a | 7B | 5 GB | Apache 2.0 | 256K | — |
| — | Magistral Small 24BMistral | n/a | 24B | 14 GB | Apache 2.0 | 128K | — |
| — | Mistral Large 3 675BMistral | n/a | 675B | 405 GB | Apache 2.0 | 256K | — |
| — | Mistral Medium 3.5 128BMistral | n/a | 128B | 74 GB | 修改版 MIT | 256K | — |
| — | Llama 4 Scout 109BLlama | n/a | 109B | 65 GB | Llama 4 Community | 10000K | — |
| — | Llama 4 Maverick 400BLlama | n/a | 400B | 240 GB | Llama 4 Community | 1000K | — |
| — | Llama 3.1 Nemotron 70BNemotron | n/a | 70B | 40 GB | Llama 3.1 Community | 128K | — |
| — | Qwen 2.5 Coder 1.5B InstructQwen | n/a | 1.5B | 1 GB | Apache 2.0 | 32K | — |
| — | Qwen 2.5 Coder 3B InstructQwen | n/a | 3B | 2 GB | Qwen Research License | 32K | — |
| — | Qwen 2.5 Coder 14B InstructQwen | n/a | 14B | 9 GB | Apache 2.0 | 131K | — |
| — | Qwen 3 30B-A3BQwen | n/a | 30B | 19 GB | Apache 2.0 | 131K | — |
| — | DeepSeek R1 Distill Qwen 1.5BDeepSeek | n/a | 1.5B | 1 GB | MIT | 131K | — |
| — | DeepSeek R1 Distill Qwen 14BDeepSeek | n/a | 14B | 9 GB | MIT | 131K | — |
| — | Phi-4 Mini Reasoning 3.8BPhi | n/a | 3.8B | 3 GB | MIT | 128K | — |
| — | Gemma 3n E2BGemma | n/a | 2B | 2 GB | Gemma | 32K | — |
| — | Gemma 3n E4BGemma | n/a | 4B | 5 GB | Gemma | 32K | — |
| — | Granite 3.3 8B InstructGranite | n/a | 8B | 5 GB | Apache 2.0 | 128K | — |
| — | Granite 4.0 H-Small 32B-A9BGranite | n/a | 32B | 19 GB | Apache 2.0 | 128K | — |
| — | Granite 4.0 H-Tiny 7B-A1BGranite | n/a | 7B | 4 GB | Apache 2.0 | 128K | — |
| — | Tülu 3 8BTulu | n/a | 8B | 6 GB | Llama 3.1 Community | 128K | — |
| — | Tülu 3 70BTulu | n/a | 70B | 40 GB | Llama 3.1 Community | 128K | — |
| — | Molmo 7B-DMolmo | n/a | 7B | 5 GB | Apache 2.0 | 4K | — |
| — | Molmo 72BMolmo | n/a | 72B | 42 GB | Apache 2.0 | 4K | — |
| — | MiniCPM-V 2.6 8BMiniCPM | n/a | 8B | 6 GB | MiniCPM模型许可 | 32K | — |
| — | MiniCPM-o 2.6 8BMiniCPM | n/a | 8B | 6 GB | MiniCPM模型许可 | 32K | — |
| — | Command R 35B v01Command | n/a | 35B | 20 GB | CC-BY-NC 4.0 | 128K | — |
| — | Aya 23 8BAya | n/a | 8B | 5 GB | CC-BY-NC 4.0 | 8K | — |
| — | Aya 23 35BAya | n/a | 35B | 20 GB | CC-BY-NC 4.0 | 8K | — |
| — | Yi Coder 9B ChatYi | n/a | 9B | 6 GB | Apache 2.0 | 128K | — |
| — | Jais 30B Chat v3Jais | n/a | 30B | 18 GB | Apache 2.0 | 8K | — |
| — | Jais Adapted 70B ChatJais | n/a | 70B | 40 GB | Apache 2.0 | 4K | — |
| — | Sarvam-M 24BSarvam | n/a | 24B | 14 GB | Apache 2.0 | 32K | — |
| — | Salamandra 7B InstructSalamandra | n/a | 7.7B | 5 GB | Apache 2.0 | 8K | — |
| — | Salamandra 40B InstructSalamandra | n/a | 40B | 24 GB | Apache 2.0 | 8K | — |
| — | EuroLLM 22B Instruct 2512EuroLLM | n/a | 22.6B | 13 GB | Apache 2.0 | 32K | — |
| — | Claire 7B 0.1Claire | n/a | 7B | 5 GB | CC-BY-NC-SA 4.0 | 2K | — |
| — | Jamba 1.5 MiniJamba | n/a | 52B | 30 GB | Jamba Open Model License | 256K | — |
| — | Hunyuan-A13B InstructHunyuan | n/a | 80B | 48 GB | Tencent Hunyuan License | 262K | — |
| — | LLaVA-OneVision 7BLLaVA | n/a | 7B | 5 GB | Apache 2.0 | 32K | — |
| — | LLaVA-OneVision 72BLLaVA | n/a | 72B | 42 GB | Apache 2.0 | 32K | — |
| — | Grok-1 (base)Grok | n/a | 314B | 188 GB | Apache 2.0 | 8K | — |
| — | gpt-oss 120Bgpt-oss | n/a | 117B | 70 GB | Apache 2.0 | 128K | — |
| — | gpt-oss 20Bgpt-oss | n/a | 21B | 13 GB | Apache 2.0 | 128K | — |
| — | Kimi K2.6Kimi | n/a | 1000B | 600 GB | 修改版 MIT | 256K | — |
| — | Qwen 3 VL 235B-A22BQwen | n/a | 235B | 142 GB | Apache 2.0 | 262K | — |
| — | Qwen 3 VL 30B-A3BQwen | n/a | 30B | 19 GB | Apache 2.0 | 262K | — |
| — | Qwen 3 VL 8BQwen | n/a | 8B | 6 GB | Apache 2.0 | 262K | — |
| — | ERNIE 4.5 300B-A47BERNIE | n/a | 300B | 180 GB | Apache 2.0 | 131K | — |
| — | ERNIE 4.5 21B-A3B ThinkingERNIE | n/a | 21B | 13 GB | Apache 2.0 | 131K | — |
| — | Ring-1TRing | n/a | 1000B | 600 GB | MIT | 131K | — |
| — | Seed-OSS 36B InstructSeed | n/a | 36B | 22 GB | Apache 2.0 | 524K | — |
| — | EXAONE 4.5 33BEXAONE | n/a | 33B | 20 GB | EXAONE AI 模型许可 | 262K | — |
| — | Nemotron Nano 3 30B-A3BNemotron | n/a | 30B | 19 GB | NVIDIA 开放模型许可证 | 1000K | — |
| — | Nemotron Nano v2 VL 12BNemotron | n/a | 12.6B | 8 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | Apertus 70BApertus | n/a | 70B | 40 GB | Apache 2.0 | 65K | — |
| — | Apertus 8BApertus | n/a | 8B | 6 GB | Apache 2.0 | 65K | — |
| — | Trinity Mini 26B-A3BTrinity | n/a | 26B | 15 GB | Apache 2.0 | 131K | — |
| — | Hunyuan Large 2.0Hunyuan | n/a | 406B | 245 GB | Tencent Hunyuan License | 262K | — |
| — | InternVL 3.5 8BInternVL | n/a | 8B | 6 GB | Apache 2.0 | 32K | — |
| — | MiMo V2 FlashMiMo | n/a | 309B | 185 GB | MIT | 128K | — |
| — | Rakuten AI 3.0Rakuten | n/a | 700B | 420 GB | Apache 2.0 | 32K | — |
| — | Kanana 2 30B-A3B ThinkingKanana | n/a | 30B | 18 GB | Apache 2.0 | 131K | — |
| — | DeepSeek-OCRDeepSeek | n/a | 3B | 2 GB | MIT | 8K | — |
| — | HunyuanOCR 1BHunyuan | n/a | 1B | 1 GB | Tencent Hunyuan License | 8K | — |
| — | Gemma 4 26B-A4B MoEGemma | n/a | 26B | 16 GB | Apache 2.0 | 128K | — |
| — | dots.llm1 Instructdots | n/a | 142B | 85 GB | MIT | 32K | — |
| — | Qwen 3 Omni 30B-A3BQwen | n/a | 30B | 19 GB | Apache 2.0 | 131K | — |
| — | Qwen 3.5 122B-A10BQwen | n/a | 122B | 73 GB | Apache 2.0 | 262K | — |
| — | Pangu Pro MoE 72BPangu | n/a | 72B | 42 GB | Pangu模型许可 | 32K | — |
| — | Qwen 3.6 27BQwen | n/a | 27B | 16 GB | Apache 2.0 | 262K | — |
| — | DeepSeek V4 Pro 1.6TDeepSeek | n/a | 1600B | 960 GB | MIT | 1000K | — |
| — | DeepSeek V4 Flash 284BDeepSeek | n/a | 284B | 170 GB | MIT | 1000K | — |
| — | Tencent Hy3 Preview 295BHunyuan | n/a | 295B | 177 GB | Tencent Hunyuan License | 256K | — |
| — | LLaDA 2.0 Uni 16BLLaDA | n/a | 16B | 18 GB | Apache 2.0 | 8K | — |
| — | MiMo V2.5 ProMiMo | n/a | 1020B | 595 GB | MIT | 1000K | — |
| — | MiMo V2.5MiMo | n/a | 310B | 180 GB | MIT | 1000K | — |
| — | Nemotron 3 Nano Omni 30B-A3BNemotron | n/a | 30B | 21 GB | NVIDIA 开放模型许可证 | 256K | — |
| — | Laguna XS.2Laguna | n/a | 33B | 19 GB | Apache 2.0 | 131K | — |
| — | Granite 4.1 30B InstructGranite | n/a | 30B | 17 GB | Apache 2.0 | 131K | — |
| — | Granite 4.1 3B InstructGranite | n/a | 3B | 2 GB | Apache 2.0 | 131K | — |
| — | Ling 2.6 1TLing | n/a | 1000B | 580 GB | MIT | 262K | — |
| — | Gemma 4 E2BGemma | n/a | 2B | 3 GB | Apache 2.0 | 128K | — |
| — | Nemotron Cascade 2 30B-A3BNemotron | n/a | 30B | 17 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | Nemotron 3 33BNemotron | n/a | 33B | 19 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | Nemotron 3 Nano 30B-A3BNemotron | n/a | 30B | 17 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | MedGemma 4BGemma | n/a | 4B | 2 GB | Gemma | 128K | — |
| — | Gemma 4 2BGemma | n/a | 2B | 1 GB | Gemma | 128K | — |
| — | Qwen 3.5 0.8BQwen | n/a | 0.8B | 1 GB | Apache 2.0 | 256K | — |
| — | MedGemma 1.5 4BGemma | n/a | 4B | 2 GB | Gemma | 128K | — |
| — | Granite 4.1Granite | n/a | 3B | 2 GB | Apache 2.0 | 128K | — |
| — | LFM2.5 Thinking 1.2BLFM | n/a | 1.2B | 1 GB | LFM Open License v1.0 | 32K | — |
| — | GLM 4.7 FlashGLM | n/a | 31B | 19 GB | MIT | 128K | — |
| — | OSINT Researcher 4BQwen3 | n/a | 4B | 2 GB | Apache 2.0 | 32K | — |
| — | LFM2 24BLFM | n/a | 24B | 14 GB | LFM Open License v1.0 | 32K | — |
| — | GLM 5 744B-A40BGLM | n/a | 744B | 432 GB | MIT | 128K | — |
| — | MiniMax M2.5 2.5BMiniMax | n/a | 2.5B | 1 GB | Apache 2.0 | 128K | — |
| — | GLM-OCR 1.1BGLM | n/a | 1.1B | 1 GB | MIT | 131K | — |
| — | Nemotron 3 Super 12BNemotron | n/a | 12B | 7 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | MiniCPM5 1B SFTMiniCPM | n/a | 1.1B | 1 GB | Apache 2.0 | 32K | — |
| — | MiniCPM5 1BMiniCPM | n/a | 1.1B | 1 GB | Apache 2.0 | 32K | — |
| — | HRM-Text 1BHRM | n/a | 1.2B | 1 GB | Apache 2.0 | 2K | — |
| — | Llama 3.1 70B LatamGPT SFTLlama | n/a | 71B | 41 GB | Llama 3.1 Community | 128K | — |
| — | Nemotron 3 UltraNemotron | n/a | 550B | 319 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | Nemotron 3 Ultra Base (BF16)Nemotron | n/a | 561B | 325 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | Nemotron 3 Ultra (BF16)Nemotron | n/a | 561B | 325 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | LFM2.5 7BLFM | n/a | 7B | 4 GB | LFM Open License v1.0 | 32K | — |
| — | Granite Code 20B (Mixed Precision)Granite | n/a | 20B | 12 GB | Apache 2.0 | 128K | — |
| — | DiffusionGemma 26B-A4B InstructGemma | n/a | 26B | 15 GB | Apache 2.0 | 128K | — |
| — | Granite 4.1 Guardian 8BGranite | n/a | 8B | 5 GB | Apache 2.0 | 128K | — |
| — | Nemotron 3.5 Content SafetyNemotron | n/a | 4.3B | 3 GB | Apache 2.0 | 128K | — |
| — | GLM 5.2 753B-A40BGLM | n/a | 753B | 437 GB | MIT | 1000K | — |
| — | Granite Embedding Multilingual R2Granite | n/a | 7B | 4 GB | Apache 2.0 | 128K | — |
| — | Nemotron TwoTower 30B-A3B BaseNemotron | n/a | 30B | 17 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | Nemotron 3 Puzzle 75B-A9BNemotron | n/a | 75B | 44 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | Qwythos 9B Claude MythosQwen | n/a | 9B | 5 GB | Apache 2.0 | 1000K | — |
| — | MiniMax M3MiniMax | n/a | 427B | 248 GB | MiniMax（自定义） | 1048K | — |
| — | MiniCPM5 1B Fable ThinkingMiniCPM | n/a | 1B | 1 GB | Apache 2.0 | 131K | — |
| — | InklingInkling | n/a | 975B | 566 GB | Apache 2.0 | 1048K | — |
| — | Kimi K2.7 CodeKimi | n/a | 1059B | 614 GB | 其他（开放权重） | 262K | — |
| — | Legal Chatbot Qwen 1.5BQwen | n/a | 1.5B | 1 GB | Apache 2.0 | 32K | — |
| — | IOL-AI Qwen3.5 9B Reasoning v2Qwen | n/a | 9.7B | 6 GB | Apache 2.0 | 262K | — |
| — | Qwen 3.5 4BQwen | n/a | 4B | 2 GB | Qwen 许可证 | 32K | — |
| — | DeepSeek V4 Flash 0731 304BDeepSeek | n/a | 304B | 176 GB | MIT | 1048K | — |
| — | LongCat Flash Lite Sparse 69B-A3BLongCat | n/a | 69B | 40 GB | MIT | 1000K | — |
| — | LFM2.5 2.6BLFM | n/a | 2.6B | 2 GB | LFM Open License v1.0 | 128K | — |
| — | Kimi K3Kimi | n/a | 2800B | 1624 GB | Kimi许可 | 1000K | — |
| — | Laguna S 2.1Laguna | n/a | 118B | 68 GB | OpenMDW 1.1 | 262K | — |
| — | Laguna XS 2.1Laguna | n/a | 33B | 19 GB | OpenMDW 1.1 | 256K | — |
| — | North Mini Code 1.0North | n/a | 30.5B | 18 GB | Apache 2.0 | 488K | — |
| — | Nemotron 3.5 Lightning 30B-A3BNemotron | n/a | 30B | 25 GB | NVIDIA 开放模型许可证 | 128K | — |
| — | Gemma 4 12BGemma | n/a | 12B | 7 GB | Apache 2.0 | 262K | — |
| — | LFM2.5 VL 3BLFM | n/a | 3B | 2 GB | LFM Open License v1.0 | 128K | — |
| — | DeepSeek V4 Pro 0813 1.7TDeepSeek | n/a | 1700B | 986 GB | MIT | 1048K | — |
| — | Qwen 3.8 27BQwen | n/a | 27B | 16 GB | Apache 2.0 | 262K | — |
| — | Nemotron 3.5 Lightning 30B-A3B (BF16)Nemotron | n/a | 30B | 25 GB | NVIDIA 开放模型许可证 | 262K | — |
| — | DeepSeek V4 Flash Coder 284B-A13B (MoEspresso V2)DeepSeek | n/a | 284B | 165 GB | MIT | 1048K | — |
| — | Qwen 3.8 27B ObliteratedQwen | n/a | 28B | 16 GB | Apache 2.0 | 262K | — |
| — | Llama 3.1 8B Reward-Hacks Inoculation (seed4)Llama | n/a | 8B | 5 GB | Apache 2.0 | 131K | — |
| — | Granite 4.2 8BGranite | n/a | 8B | 5 GB | Apache 2.0 | 128K | — |
| — | LFM2.5 DSparkLFM | n/a | 7B | 4 GB | LFM Open License v1.0 | 32K | — |
| — | S1-miniQwen | n/a | 0.6B | 0 GB | 其他（开放权重） | 40K | — |
| — | Qwen3.8 Flash Next 125B-A6BQwen | n/a | 125B | 72 GB | 其他（开放权重） | 256K | — |
| — | GLM 5.3 Flash 320B-A18BGLM | n/a | 320B | 186 GB | MIT | 128K | — |
| — | IBM Granite Code 7B DPOGranite | n/a | 7B | 4 GB | Apache 2.0 | 128K | — |
| — | IBM Granite Code 8B InstructGranite | n/a | 8B | 5 GB | Apache 2.0 | 128K | — |
| — | OLMo 3 7B Think (SFT)OLMo | n/a | 7B | 4 GB | Apache 2.0 | 16K | — |
| — | GLM 5.3 7BGLM | n/a | 7B | 4 GB | MIT | 128K | — |
| — | IBM Granite Code 7B v3Granite | n/a | 7B | 4 GB | Apache 2.0 | 128K | — |
| — | Activity Generation 0.5B (Qwen)Qwen | n/a | 0.5B | 0 GB | Apache 2.0 | 32K | — |
| — | Qwen3 4B NemotronIF Reasoning SFTQwen | n/a | 4B | 2 GB | Apache 2.0 | 32K | — |
| — | OLMo 7B CPT Merged (step 750)OLMo | n/a | 7B | 4 GB | Apache 2.0 | 65K | — |
| — | Granite Time Series PatchTST R2Granite | n/a | 0.4B | 0 GB | Apache 2.0 | 8K | — |
| — | Qwen 3.8 27B Obliterated Mythos AgenticQwen | n/a | 28B | 16 GB | Apache 2.0 | 262K | — |
| — | DeepSeek V4.1 Flash 552BDeepSeek | n/a | 552B | 320 GB | DeepSeek | 1000K | — |
| — | IBM Granite Code 8BGranite | n/a | 8B | 5 GB | Apache 2.0 | 128K | — |
| — | IBM Granite 4.2Granite | n/a | 4.2B | 2 GB | Apache 2.0 | 128K | — |
| — | Shieldstral 3BMistral | n/a | 3B | 2 GB | Apache 2.0 | 32K | — |
| — | EmbeddingGemma 2Gemma | n/a | 0.74B | 0 GB | Apache 2.0 | 8K | — |
| — | Mistral Large 4Mistral | n/a | 1000B | 580 GB | 其他（开放权重） | 32K | — |
| — | Qwen 3 27BQwen | n/a | 27B | 16 GB | Apache 2.0 | 32K | — |
该排行榜的评分标准
每个模型都采用加权综合评分，得分范围为 0 至 100： 45% 质量 （MMLU 基准测试）， 效率占 25% （每 GB 显存对应的得分——当模型运行在 votre 硬件), 15% 硬件适配性 （能否装入 8/12/16/24 GB 显存中？）， 许可证占10% （是否允许自由商用？）以及 5% 上下文窗口. 每个子指标均基于 34 个跟踪模型进行 min-max 归一化处理，因此得分 以该集合为参照的 并在每次更新时重新计算。VRAM 数值和效率指标均假定采用量化格式 Q4_K_M。基准测试分数来自各模型已公布的结果（参见其 目录中的模型详情页）；没有参考基准测试成绩的模型列为 n/a 并排在最后，不对其分数进行估算。
| 指标 | 权重 | 衡量的内容 |
|---|---|---|
| 质量 | 45 % | MMLU基准测试（标准化） |
| 效率 | 25 % | 每 GB 显存的基准测试表现 |
| 可访问性 | 15 % | 可装入 8/12/16/24 GB 显存 |
| 许可证 | 10 % | 商业使用自由 |
| 上下文 | 5 % | 可用上下文窗口 |
常见问题
该排行榜是如何计算的？
每个模型都采用加权综合评分，得分范围为 0 至 100： 45% 质量 (MMLU基准测试), 效率占 25% (按每 GB 显存计算的得分) 15% 硬件适配性 （能否装入 8/12/16/24 GB 显存中？）， 许可证占10% （是否允许自由商用？）以及 5% 上下文。每个子指标均以目录中经过基准测试的 34 个模型为基础，进行最小值—最大值归一化。
为什么某些模型显示「n/a」？
我们目录中没有已公布 MMLU 分数的模型会列在表格末尾，分数标为“n/a”，绝不估算或插补——诚实比排名的完整性更重要。
这些 VRAM 数值是否以某种特定的量化方式为前提？
是的，最小显存需求和效率计算均假设采用 Q4_K_M 量化，这是大众本地推理中最常用的量化方式。
这些数据从哪里来？
来自 QuelLLM.fr 的公开目录（project/data/models.js + details.js），我们的模型介绍、比较工具和按用途划分的排名也使用这一数据源。