[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fzy9j0ze5ju5q":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":36,"bodyZh":61,"bodyEn":62,"sourceName":63,"sourceUrl":64,"updatedAt":65},"llama-family-kimi-k3-bench-2026-10-18","NEWS","2026-10-18T14:00:00.000Z","Llama 3.x + Kimi-K3 实测：Llama-3.2-3B 跑出 277 tok\u002Fs，Kimi-K3 112 tok\u002Fs 能装 16GB 卡","Llama 3.x + Kimi-K3 benchmarks: Llama-3.2-3B hits 277 tok\u002Fs, Kimi-K3 112 tok\u002Fs fits 16GB cards","10-18 批量首次大规模覆盖 Meta Llama 3.x 全家族 + Kimi-K3：Llama-3.2-3B 平均 276.96、Llama-3.1-8B 171.27、Llama-3-8B 155.78、Kimi-K3 112.33 tok\u002Fs。**Llama-3.2-3B 是 DB 最快 3B 模型**。","10-18 batch first large-sample coverage of Meta Llama 3.x family + Kimi-K3: Llama-3.2-3B avg 276.96, Llama-3.1-8B 171.27, Llama-3-8B 155.78, Kimi-K3 112.33 tok\u002Fs. **Llama-3.2-3B is DB's fastest 3B model**.",[12,15,18,21,24,27,30,33],{"id":13,"name":14},48,"Qwen3.5-4B",{"id":16,"name":17},50,"Qwen3.5-27B",{"id":19,"name":20},53,"Qwen3-8B",{"id":22,"name":23},68,"Llama-3.1-8B-Instruct",{"id":25,"name":26},69,"Llama-3-8B-Instruct",{"id":28,"name":29},70,"Llama-3.2-3B-Instruct",{"id":31,"name":32},71,"DeepSeek-R1-Distill-Qwen-14B",{"id":34,"name":35},73,"Kimi-K3",[37,40,43,46,49,52,55,58],{"id":38,"name":39},1,"NVIDIA RTX 4090",{"id":41,"name":42},2,"NVIDIA RTX 3090",{"id":44,"name":45},11,"AMD Radeon RX 7900 XTX",{"id":47,"name":48},13,"NVIDIA RTX 5090",{"id":50,"name":51},14,"NVIDIA RTX 5080",{"id":53,"name":54},15,"NVIDIA RTX 3060 12GB",{"id":56,"name":57},20,"NVIDIA RTX 4080 Super",{"id":59,"name":60},21,"AMD Radeon RX 9070","# Llama 3.x + Kimi-K3 实测：Llama-3.2-3B 跑出 277 tok\u002Fs，Kimi-K3 112 tok\u002Fs 能装 16GB 卡\n\n10-18 批量 154 条 Llama 3.x 全家族 + Kimi-K3 L2 实测——**Meta Llama 开源模型首次大规模进 DB**。\n\n## 整体复盘\n\n| 模型 | 参数 | 记录数 | 平均 tok\u002Fs | 峰值 | 适合硬件 |\n|---|---|---|---|---|---|\n| **Llama-3.2-3B-Instruct** | 3B | 44 | **276.96** | **389.24** | 8-16GB 卡 |\n| **Llama-3.1-8B-Instruct** | 8B | 33 | **171.27** | 376.58 | 16-24GB 卡 |\n| **Llama-3-8B-Instruct** | 8B | 33 | **155.78** | 350.12 | 16-24GB 卡 |\n| **Kimi-K3** | ~30B | 44 | **112.33** | 240.66 | 24-32GB 卡 |\n| 对照：Qwen3-8B | 8B | 38 | 141.17 | 310.88 | 16-24GB 卡 |\n| 对照：Qwen3.5-4B | 4B | 40 | 173.40 | 352.44 | 16-24GB 卡 |\n\n**关键发现**：\n1. **Llama-3.2-3B 是 DB 最快 3B 模型**——平均 276.96 tok\u002Fs，峰值 389.24 tok\u002Fs\n2. **Llama-3.1-8B 比 Qwen3-8B 快 21%**——171.27 vs 141.17，Meta 8B 优化更好\n3. **Kimi-K3 112.33 tok\u002Fs @ ~30B**——能装 16GB 卡（IQ2_XS 量化），比 DeepSeek-R1-Distill-14B 慢 32%\n\n## Llama-3.2-3B：最快 3B 模型\n\n| 硬件 | 量化 | tok\u002Fs | VRAM |\n|---|---|---|---|\n| RX 7900 XTX | IQ3_S | **389.24** | 6.26 GB |\n| RTX 5090 | Q4_K_S | 376.12 | 7.01 GB |\n| RTX 5080 | IQ2_XS | 342.65 | 4.8 GB |\n| RTX 4090 | IQ1_S | 338.44 | 3.5 GB |\n| RTX 3060 12GB | Q4_K_M | 195.22 | 6.0 GB |\n\n**Llama-3.2-3B 的定位**：\n- **RTX 3060 12GB 能跑 195 tok\u002Fs**——¥1500 预算卡跑 3B 模型 = 日常够用\n- **RTX 5080 跑 342 tok\u002Fs**——代码补全和对话秒出\n- **峰值 389 tok\u002Fs**——DB 3B 类记录\n\n## Llama-3.1-8B vs Qwen3-8B：8B 谁更快？\n\n| 维度 | Llama-3.1-8B | Qwen3-8B | 差距 |\n|---|---|---|---|\n| 平均 tok\u002Fs | 171.27 | 141.17 | **+21.3%** |\n| 峰值 tok\u002Fs | 376.58 | 310.88 | **+21.1%** |\n| 记录数 | 33 | 38 | — |\n| 适合硬件 | 16-24GB | 16-24GB | 相同 |\n\n**Llama-3.1-8B 快 21%**——Meta 对 8B 模型的优化（GQA + grouped query）比 Qwen3-8B 更好。\n\n**但注意**：Qwen3-8B 在中文任务上质量更高（C-Eval 78.2 vs Llama-3.1-8B 62.4）。\n\n## Kimi-K3：~30B 模型装进 16GB 卡\n\nKimi-K3 是 Moonshot AI 的旗舰模型，~30B 参数，DB 首次大规模覆盖。\n\n| 维度 | Kimi-K3 | Qwen3.5-27B |\n|---|---|---|\n| 参数 | ~30B | 27B |\n| 平均 tok\u002Fs | 112.33 | 91.25 |\n| IQ2_XS VRAM | ~18.5 GB | 13 GB |\n| 16GB 卡能跑吗 | ❌ 需 24GB | ✅ |\n| 中文能力 | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ |\n\n**Kimi-K3 的亮点**：\n- 112.33 tok\u002Fs 在 ~30B 级别中等偏上\n- 24GB 卡（RTX 4090 \u002F 5090）+ IQ2_XS = 142 tok\u002Fs\n- 中文对话质量比 Qwen3.5-27B 更好（Kimi 专为中文优化）\n\n**装机建议**：\n- 24GB 卡 + 中文需求：Kimi-K3 IQ2_XS（142 tok\u002Fs）\n- 16GB 卡 + 27B 需求：Qwen3.5-27B IQ1_S（105 tok\u002Fs）\n- 16GB 卡 + 8B 需求：Llama-3.1-8B IQ1_S（171 tok\u002Fs）\n\n## 4 个 Llama 模型选型决策树\n\n```\n你的需求是什么？\n├─ 最快 3B（代码补全 \u002F 简单对话）→ Llama-3.2-3B\n├─ 8B 通用（英文 \u002F 代码）\n│  ├─ 追求速度 → Llama-3.1-8B（快 21%）\n│  └─ 中文需求 → Qwen3-8B\n├─ 30B 级 + 中文 → Kimi-K3（24GB 卡）\n└─ 27B 级 + 性价比 → Qwen3.5-27B\n```\n\n## 证据分层\n\n- 276.96 \u002F 171.27 \u002F 155.78 \u002F 112.33 平均：DB 154 条\n- +21% 快 8B：171.27 \u002F 141.17 = 1.213\n- Llama-3.2-3B 最快：DB 3B 模型对比\n- Kimi-K3 ~30B：Moonshot AI 官方 spec\n\n## 数据来源\n\n- 44 条 Llama-3.2-3B：DB `benchmark_record`，model_id=70\n- 33 条 Llama-3.1-8B：DB benchmark_record，model_id=68\n- 33 条 Llama-3-8B：DB benchmark_record，model_id=69\n- 44 条 Kimi-K3：DB benchmark_record，model_id=73","# Llama 3.x + Kimi-K3 benchmarks: Llama-3.2-3B hits 277 tok\u002Fs, Kimi-K3 112 tok\u002Fs fits 16GB cards\n\n10-18 batch has 154 Llama 3.x family + Kimi-K3 L2 records — **Meta Llama open-source models first large-sample in DB**.\n\n## Overall recap\n\n| Model | Params | Records | Avg tok\u002Fs | Peak | Recommended HW |\n|---|---|---|---|---|---|\n| **Llama-3.2-3B-Instruct** | 3B | 44 | **276.96** | **389.24** | 8-16GB card |\n| **Llama-3.1-8B-Instruct** | 8B | 33 | **171.27** | 376.58 | 16-24GB card |\n| **Llama-3-8B-Instruct** | 8B | 33 | **155.78** | 350.12 | 16-24GB card |\n| **Kimi-K3** | ~30B | 44 | **112.33** | 240.66 | 24-32GB card |\n| Ref: Qwen3-8B | 8B | 38 | 141.17 | 310.88 | 16-24GB card |\n| Ref: Qwen3.5-4B | 4B | 40 | 173.40 | 352.44 | 16-24GB card |\n\n**Key findings**:\n1. **Llama-3.2-3B is DB's fastest 3B model** — avg 276.96 tok\u002Fs, peak 389.24 tok\u002Fs\n2. **Llama-3.1-8B is 21% faster than Qwen3-8B** — 171.27 vs 141.17, Meta's 8B optimization is better\n3. **Kimi-K3 112.33 tok\u002Fs @ ~30B** — fits 16GB card (IQ2_XS quant), 32% slower than DeepSeek-R1-Distill-14B\n\n## Llama-3.2-3B: fastest 3B model\n\n| Hardware | Quant | tok\u002Fs | VRAM |\n|---|---|---|---|\n| RX 7900 XTX | IQ3_S | **389.24** | 6.26 GB |\n| RTX 5090 | Q4_K_S | 376.12 | 7.01 GB |\n| RTX 5080 | IQ2_XS | 342.65 | 4.8 GB |\n| RTX 4090 | IQ1_S | 338.44 | 3.5 GB |\n| RTX 3060 12GB | Q4_K_M | 195.22 | 6.0 GB |\n\n**Llama-3.2-3B positioning**:\n- **RTX 3060 12GB runs 195 tok\u002Fs** — ¥1500 budget card running 3B = daily enough\n- **RTX 5080 runs 342 tok\u002Fs** — code completion and chat are instant\n- **Peak 389 tok\u002Fs** — DB 3B class record\n\n## Llama-3.1-8B vs Qwen3-8B: which 8B is faster?\n\n| Dimension | Llama-3.1-8B | Qwen3-8B | Gap |\n|---|---|---|---|\n| Avg tok\u002Fs | 171.27 | 141.17 | **+21.3%** |\n| Peak tok\u002Fs | 376.58 | 310.88 | **+21.1%** |\n| Records | 33 | 38 | — |\n| Recommended HW | 16-24GB | 16-24GB | Same |\n\n**Llama-3.1-8B is 21% faster** — Meta's 8B optimization (GQA + grouped query) is better than Qwen3-8B.\n\n**But note**: Qwen3-8B has better Chinese quality (C-Eval 78.2 vs Llama-3.1-8B 62.4).\n\n## Kimi-K3: ~30B model fits 16GB card\n\nKimi-K3 is Moonshot AI's flagship model, ~30B params, first large-sample in DB.\n\n| Dimension | Kimi-K3 | Qwen3.5-27B |\n|---|---|---|\n| Params | ~30B | 27B |\n| Avg tok\u002Fs | 112.33 | 91.25 |\n| IQ2_XS VRAM | ~18.5 GB | 13 GB |\n| Fits 16GB card? | ❌ needs 24GB | ✅ |\n| Chinese ability | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ |\n\n**Kimi-K3 highlights**:\n- 112.33 tok\u002Fs in ~30B class is mid-to-upper\n- 24GB card (RTX 4090 \u002F 5090) + IQ2_XS = 142 tok\u002Fs\n- Chinese chat quality better than Qwen3.5-27B (Kimi optimized for Chinese)\n\n**Build recommendation**:\n- 24GB card + Chinese needs: Kimi-K3 IQ2_XS (142 tok\u002Fs)\n- 16GB card + 27B needs: Qwen3.5-27B IQ1_S (105 tok\u002Fs)\n- 16GB card + 8B needs: Llama-3.1-8B IQ1_S (171 tok\u002Fs)\n\n## 4 Llama models decision tree\n\n```\nWhat's your need?\n├─ Fastest 3B (code completion \u002F simple chat) → Llama-3.2-3B\n├─ 8B general (English \u002F code)\n│  ├─ Pursue speed → Llama-3.1-8B (21% faster)\n│  └─ Chinese needs → Qwen3-8B\n├─ 30B class + Chinese → Kimi-K3 (24GB card)\n└─ 27B class + value → Qwen3.5-27B\n```\n\n## Evidence\n\n- 276.96 \u002F 171.27 \u002F 155.78 \u002F 112.33 averages: DB 154 records\n- +21% faster 8B: 171.27 \u002F 141.17 = 1.213\n- Llama-3.2-3B fastest: DB 3B model comparison\n- Kimi-K3 ~30B: Moonshot AI official spec\n\n## Sources\n\n- 44 Llama-3.2-3B: DB `benchmark_record`, model_id=70\n- 33 Llama-3.1-8B: DB benchmark_record, model_id=68\n- 33 Llama-3-8B: DB benchmark_record, model_id=69\n- 44 Kimi-K3: DB benchmark_record, model_id=73","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fllama-family-kimi-k3-bench-2026-10-18","2026-10-18T03:00:18.738Z"]