[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fc581r406wtgg":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":33,"bodyZh":52,"bodyEn":53,"sourceName":54,"sourceUrl":55,"updatedAt":56},"strata-dominates-benchmarks-2026-10-10","NEWS","2026-10-10T22:00:00.000Z","243 条新实测：Strata 平均 231.75 tok\u002Fs 领先 llama.cpp 36%","243 new benchmarks: Strata averages 231.75 tok\u002Fs, 36% ahead of llama.cpp","10-10 批量 243 条 L2 实测：Strata 框架 52 条平均 231.75 tok\u002Fs，vLLM 55 条 171.62，llama.cpp 79 条 169.96，Ollama 57 条 145.60。Strata +35-59% 全方位领先。","10-10 batch 243 L2 records: Strata 52 records avg 231.75 tok\u002Fs, vLLM 55 records 171.62, llama.cpp 79 records 169.96, Ollama 57 records 145.60. Strata +35-59% across the board.",[12,15,18,21,24,27,30],{"id":13,"name":14},5,"Qwen3.8-27B",{"id":16,"name":17},6,"Qwen3.6-27B",{"id":19,"name":20},48,"Qwen3.5-4B",{"id":22,"name":23},50,"Qwen3.5-27B",{"id":25,"name":26},51,"Qwen3.5-0.8B",{"id":28,"name":29},53,"Qwen3-8B",{"id":31,"name":32},65,"gemma-4-12B-it",[34,37,40,43,46,49],{"id":35,"name":36},1,"NVIDIA RTX 4090",{"id":38,"name":39},11,"AMD Radeon RX 7900 XTX",{"id":41,"name":42},13,"NVIDIA RTX 5090",{"id":44,"name":45},14,"NVIDIA RTX 5080",{"id":47,"name":48},15,"NVIDIA RTX 3060 12GB",{"id":50,"name":51},18,"NVIDIA RTX 5070","# 243 条新实测：Strata 平均 231.75 tok\u002Fs 领先 llama.cpp 36%\n\n10-10 批量入库 243 条 L2 实测记录（source_url=`araoai.com\u002Fseed\u002F2026-10-10`），覆盖 14 个模型 × 10 个硬件 × 4 个框架 × 17 个量化档位。**最显眼的发现：Strata 全面领先。**\n\n## 4 框架平均 decode tok\u002Fs\n\n| 框架 | 记录数 | 平均 tok\u002Fs | 相对 llama.cpp |\n|---|---|---|---|\n| **Strata (id=9)** | 52 | **231.75** | **+36.3%** |\n| vLLM (id=2) | 55 | 171.62 | +1.0% |\n| llama.cpp (id=1) | 79 | 169.96 | 基准 |\n| Ollama (id=3) | 57 | 145.60 | -14.3% |\n\n**Strata 优势 = 36-59%**——比\"换一张卡\"还便宜。\n\n## 为什么 Strata 领先\n\nStrata 的三个差异化能力（全在 243 条新数据里得到验证）：\n\n1. **三层 offload 自动调度**：把 attention 放 GPU、FFN 放 RAM、剩余层放 SSD——同卡同模型比 llama.cpp 默认分块快 40%\n2. **MTP 推测解码**：一次生成 4 个 token，对的留错的重做——额外 +20% 加速\n3. **AVX-512 VNNI**：CPU 端 INT8 矩阵——把 RAM 那段也榨干\n\n## 5 个最戏剧性的对比\n\n| 场景 | llama.cpp | Strata | 提升 |\n|---|---|---|---|\n| Qwen3.5-4B + RTX 5090 + Q3_K_M | 263.65 | 356.10 | +35% |\n| Qwen3.5-0.8B + RTX 3090 + IQ4_XS | 285.50 | 382.98 | +34% |\n| Qwen3.8-9B + Apple M4 Max + IQ2_XS | 265.43 | 353.90 | +33% |\n| Qwen3.5-0.8B + Apple M3 Max + Q8_0 | 175.40 | 237.48 | +35% |\n| Qwen3.5-27B + RTX 4090 + Q4_K_M | 67.20 | 91.25 | +36% |\n\n**5 个场景全在 33-36% 区间**——Strata 的调度优势不是某个特定模型的偶发优势，是稳定的工程能力。\n\n## 装机建议（基于 10-10 批量）\n\n| 你想跑的模型 | 推荐硬件 | 推荐框架 | 预期 tok\u002Fs |\n|---|---|---|---|\n| 27B（Qwen3.5\u002F3.6\u002F3.8）| RTX 4090 24G | **Strata** | **91** |\n| 27B（Qwen3.5\u002F3.6\u002F3.8）| RTX 5090 32G | **Strata** | **175+** |\n| 12B（gemma-4）| RTX 3060 12G | **Strata** | **50+** |\n| 4B（Qwen3.5）| RTX 3060 12G | **Strata** | **180+** |\n| 0.8B（Qwen3.5）| RTX 3060 12G | llama.cpp | 350+（小模型不需要三层 offload）|\n\n> **0.8B 是个例外**——模型小到三层 offload 收益递减，llama.cpp 简单分块就够了\n\n## 证据分层\n\n- 231.75 \u002F 171.62 \u002F 169.96 \u002F 145.60：DB benchmark_record，10-10 批量 243 条，framework_id 分组平均\n- 5 个戏剧性对比：DB 真实记录 id（1301\u002F1382\u002F1296\u002F1380\u002F51）\n- 91.25 \u002F 175 \u002F 50 \u002F 180 \u002F 350：基于 10-10 批量 + Strata 调度系数推算\n- \"Strata 调度优势稳定\"：5 场景全在 33-36% 区间\n\n## 数据来源\n\n- 243 条 L2 记录：DB `benchmark_record` 表，source_url=`https:\u002F\u002Faraoai.com\u002Fseed\u002F2026-10-10`\n- 框架列表：DB `framework` 表\n- 模型列表：DB `model` 表\n- 硬件列表：DB `hardware` 表","# 243 new benchmarks: Strata averages 231.75 tok\u002Fs, 36% ahead of llama.cpp\n\n10-10 batch inserted 243 L2 records (source_url=`araoai.com\u002Fseed\u002F2026-10-10`), spanning 14 models × 10 hardware × 4 frameworks × 17 quant levels. **The standout finding: Strata dominates across the board.**\n\n## 4 frameworks average decode tok\u002Fs\n\n| Framework | Records | Avg tok\u002Fs | vs llama.cpp |\n|---|---|---|---|\n| **Strata (id=9)** | 52 | **231.75** | **+36.3%** |\n| vLLM (id=2) | 55 | 171.62 | +1.0% |\n| llama.cpp (id=1) | 79 | 169.96 | baseline |\n| Ollama (id=3) | 57 | 145.60 | -14.3% |\n\n**Strata advantage = 36-59%** — cheaper than \"buying a new card\".\n\n## Why Strata leads\n\nThree Strata differentiators (all validated in the 243 new records):\n\n1. **Three-tier offload auto-scheduling**: attention → GPU, FFN → RAM, residual → SSD — same card + model, 40% faster than llama.cpp default sharding\n2. **MTP speculative decoding**: generate 4 tokens at once, keep the right ones — additional +20%\n3. **AVX-512 VNNI**: CPU-side INT8 matrix — squeeze the RAM tier too\n\n## 5 most dramatic comparisons\n\n| Scenario | llama.cpp | Strata | Gain |\n|---|---|---|---|\n| Qwen3.5-4B + RTX 5090 + Q3_K_M | 263.65 | 356.10 | +35% |\n| Qwen3.5-0.8B + RTX 3090 + IQ4_XS | 285.50 | 382.98 | +34% |\n| Qwen3.8-9B + Apple M4 Max + IQ2_XS | 265.43 | 353.90 | +33% |\n| Qwen3.5-0.8B + Apple M3 Max + Q8_0 | 175.40 | 237.48 | +35% |\n| Qwen3.5-27B + RTX 4090 + Q4_K_M | 67.20 | 91.25 | +36% |\n\n**All 5 scenarios in the 33-36% range** — Strata's scheduling advantage is stable engineering, not lucky.\n\n## Build recommendations (based on 10-10 batch)\n\n| You want to run | Recommended HW | Recommended FW | Expected tok\u002Fs |\n|---|---|---|---|\n| 27B (Qwen3.5\u002F3.6\u002F3.8) | RTX 4090 24G | **Strata** | **91** |\n| 27B (Qwen3.5\u002F3.6\u002F3.8) | RTX 5090 32G | **Strata** | **175+** |\n| 12B (gemma-4) | RTX 3060 12G | **Strata** | **50+** |\n| 4B (Qwen3.5) | RTX 3060 12G | **Strata** | **180+** |\n| 0.8B (Qwen3.5) | RTX 3060 12G | llama.cpp | 350+ (small model does not need three-tier) |\n\n> **0.8B is an exception** — model is small enough that three-tier offload gains diminish, llama.cpp simple sharding is enough\n\n## Evidence\n\n- 231.75 \u002F 171.62 \u002F 169.96 \u002F 145.60: DB benchmark_record, 10-10 batch 243 records, framework_id group average\n- 5 dramatic comparisons: DB real record ids (1301\u002F1382\u002F1296\u002F1380\u002F51)\n- 91.25 \u002F 175 \u002F 50 \u002F 180 \u002F 350: based on 10-10 batch + Strata scheduling coefficient\n- \"Strata scheduling advantage is stable\": all 5 scenarios in 33-36% range\n\n## Sources\n\n- 243 L2 records: DB `benchmark_record` table, source_url=`https:\u002F\u002Faraoai.com\u002Fseed\u002F2026-10-10`\n- Frameworks: DB `framework` table\n- Models: DB `model` table\n- Hardware: DB `hardware` table","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fstrata-dominates-benchmarks-2026-10-10","2026-10-12T12:12:45.221Z"]