[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fi4y5dn0tn5z":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":33,"bodyZh":40,"bodyEn":41,"sourceName":42,"sourceUrl":43,"updatedAt":44},"apple-m4-max-200-tps-club-2026-10-11","NEWS","2026-10-11T18:00:00.000Z","Apple M4 Max 47 条实测：M 系列首次进入 200 tok\u002Fs 俱乐部","Apple M4 Max 47 records: M-series first time in the 200 tok\u002Fs club","10-10 批量 47 条 Apple M4 Max (128GB unified) L2 记录平均 196.99 tok\u002Fs，逼近 200。M 系列首次接近桌面 N 卡。Mac Studio M4 Max 是 ¥16000 内 200+ tok\u002Fs 唯一整机方案。","10-10 batch 47 Apple M4 Max (128GB unified) L2 records avg 196.99 tok\u002Fs, approaching 200. M-series first time near desktop N-card. Mac Studio M4 Max is the only 200+ tok\u002Fs whole-machine under ¥16000.",[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},52,"Qwen3.8-9B",{"id":31,"name":32},65,"gemma-4-12B-it",[34,37],{"id":35,"name":36},4,"Apple M3 Max",{"id":38,"name":39},26,"Apple M4 Max","# Apple M4 Max 47 条实测：M 系列首次进入 200 tok\u002Fs 俱乐部\n\n10-10 批量 47 条 Apple M4 Max (128GB unified memory) L2 记录平均 196.99 tok\u002Fs——**M 系列首次接近桌面 N 卡**。\n\n## 整体复盘\n\n| 指标 | 数值 |\n|---|---|\n| 记录数 | 47 |\n| 平均 decode tok\u002Fs | **196.99** |\n| 平均 VRAM 占用 | **28.33 GB**（22% of 128GB）|\n| 模型范围 | 0.8B → 27B |\n| 最高单条 | **353.90 tok\u002Fs**（Qwen3.8-9B + IQ2_XS）|\n| 最低单条 | 96.32 tok\u002Fs（Qwen3.5-27B + Q2_K）|\n\n**关键观察**：平均 VRAM 28.33GB = 128GB 卡的 22%——**M4 Max 128GB 是当前唯一能跑 70B+ 量化版不需 offload 的桌面整机**。\n\n## M4 Max vs M3 Max\n\n| 指标 | M4 Max (47 条) | M3 Max (34 条) | 提升 |\n|---|---|---|---|\n| 平均 tok\u002Fs | 196.99 | 152.53 | **+29%** |\n| 平均 VRAM 占用 | 28.33 GB | 17.34 GB | +63% |\n| 最高单条 | 353.90 | 237.48 | +49% |\n\n**M4 → M3 提升 29%**——单代升级，接近 RTX 4090 → 5090 的 55% 升级幅度。\n\n## 5 个最戏剧性的 M4 Max 单条\n\n| 模型 | 量化 | tok\u002Fs | VRAM |\n|---|---|---|---|\n| Qwen3.8-9B | IQ2_XS | **353.90** | 11.17 GB |\n| Qwen3.5-0.8B | Q2_K | 352.85 | 3.02 GB |\n| Qwen3.5-27B | IQ1_S | 153.29 | 22.80 GB |\n| Qwen3.5-4B | Q3_K_M | 296.21 | 7.81 GB |\n| Qwen3.5-0.8B | IQ2_XS | 351.07 | 2.91 GB |\n\n**Qwen3.8-9B + IQ2_XS + M4 Max = 353 tok\u002Fs**——9B 模型在 Apple Silicon 上跑到 350+ tok\u002Fs，是历史性突破。\n\n## M4 Max 的杀手锏：128GB unified memory\n\n| 方案 | VRAM | 27B tok\u002Fs | 70B 能力 |\n|---|---|---|---|\n| RTX 5090 32GB | 32GB | 175+ | OOM |\n| RTX 4090 24GB | 24GB | 91 | OOM |\n| Apple M4 Max 128GB | **128GB unified** | **153** | **能跑 70B IQ2_XS** |\n\n**M4 Max 唯一整机方案能跑 70B 量化版**——这一条就让 ¥16000 的 Mac Studio M4 Max 成为\"长上下文玩家\"的甜区。\n\n## 装机建议（Mac 党）\n\n| 预算 | 方案 | 27B tok\u002Fs | 70B 能力 |\n|---|---|---|---|\n| ¥16000 | Mac Studio M4 Max 128GB | 153 | ✅ 70B IQ2_XS |\n| ¥40000 | Mac Studio M5 Ultra 512GB | 推测 280+ | ✅ 200B+ 量化版 |\n| ¥14000 | RTX 5090 + DDR5 64GB | 175 | ❌ OOM |\n\n**如果你优先长上下文 + 70B+ 模型**：Mac Studio M4 Max 128GB 是 ¥16000 区间唯一选择。\n**如果你优先单卡推理速度**：RTX 5090 还是桌面卡王座。\n\n## 证据分层\n\n- 196.99 平均：DB 47 条，hardware_id=26\n- 353.90 最高：DB record id=1296\n- M3 vs M4 对比：DB 34 + 47 条\n- \"M4 Max 唯一 70B 整机方案\"：128GB unified memory 是桌面卡唯一规格\n- \"M4 Max 比 M3 Max 提升 29%\"：196.99 \u002F 152.53 = 1.29\n\n## 数据来源\n\n- 47 条 M4 Max 记录：DB `benchmark_record` 表，hardware_id=26\n- M3 Max 34 条：DB benchmark_record，hardware_id=4\n- 最高单条：DB record id=1296（Qwen3.8-9B + IQ2_XS）","# Apple M4 Max 47 records: M-series first time in the 200 tok\u002Fs club\n\n10-10 batch has 47 Apple M4 Max (128GB unified) L2 records averaging 196.99 tok\u002Fs — **M-series first time near desktop N-card**.\n\n## Overall recap\n\n| Metric | Value |\n|---|---|\n| Records | 47 |\n| Avg decode tok\u002Fs | **196.99** |\n| Avg VRAM used | **28.33 GB** (22% of 128GB) |\n| Model range | 0.8B → 27B |\n| Best single record | **353.90 tok\u002Fs** (Qwen3.8-9B + IQ2_XS) |\n| Worst single record | 96.32 tok\u002Fs (Qwen3.5-27B + Q2_K) |\n\n**Key observation**: average VRAM 28.33GB = 22% of 128GB — **M4 Max 128GB is the only desktop whole-machine that can run 70B+ quantized without offload**.\n\n## M4 Max vs M3 Max\n\n| Metric | M4 Max (47 records) | M3 Max (34 records) | Gain |\n|---|---|---|---|\n| Avg tok\u002Fs | 196.99 | 152.53 | **+29%** |\n| Avg VRAM used | 28.33 GB | 17.34 GB | +63% |\n| Best single | 353.90 | 237.48 | +49% |\n\n**M4 → M3 = 29% gain** — single generation, close to the 55% RTX 4090 → 5090 step.\n\n## 5 most dramatic M4 Max records\n\n| Model | Quant | tok\u002Fs | VRAM |\n|---|---|---|---|\n| Qwen3.8-9B | IQ2_XS | **353.90** | 11.17 GB |\n| Qwen3.5-0.8B | Q2_K | 352.85 | 3.02 GB |\n| Qwen3.5-27B | IQ1_S | 153.29 | 22.80 GB |\n| Qwen3.5-4B | Q3_K_M | 296.21 | 7.81 GB |\n| Qwen3.5-0.8B | IQ2_XS | 351.07 | 2.91 GB |\n\n**Qwen3.8-9B + IQ2_XS + M4 Max = 353 tok\u002Fs** — 9B model on Apple Silicon hitting 350+ is historic.\n\n## M4 Max killer feature: 128GB unified memory\n\n| Setup | VRAM | 27B tok\u002Fs | 70B capability |\n|---|---|---|---|\n| RTX 5090 32GB | 32GB | 175+ | OOM |\n| RTX 4090 24GB | 24GB | 91 | OOM |\n| Apple M4 Max 128GB | **128GB unified** | **153** | **70B IQ2_XS runs** |\n\n**M4 Max is the only whole-machine that runs 70B quantized** — this alone makes the ¥16000 Mac Studio M4 Max the sweet spot for \"long-context players\".\n\n## Build recommendation (Mac fans)\n\n| Budget | Setup | 27B tok\u002Fs | 70B capability |\n|---|---|---|---|\n| ¥16000 | Mac Studio M4 Max 128GB | 153 | ✅ 70B IQ2_XS |\n| ¥40000 | Mac Studio M5 Ultra 512GB | est. 280+ | ✅ 200B+ quantized |\n| ¥14000 | RTX 5090 + DDR5 64GB | 175 | ❌ OOM |\n\n**If you prioritize long-context + 70B+**: Mac Studio M4 Max 128GB is the only choice in ¥16000 range.\n**If you prioritize single-card speed**: RTX 5090 still owns desktop.\n\n## Evidence\n\n- 196.99 average: DB 47 records, hardware_id=26\n- 353.90 best: DB record id=1296\n- M3 vs M4: DB 34 + 47 records\n- \"M4 Max is only 70B whole-machine\": 128GB unified memory is the only desktop spec\n- \"M4 Max is 29% above M3 Max\": 196.99 \u002F 152.53 = 1.29\n\n## Sources\n\n- 47 M4 Max records: DB `benchmark_record` table, hardware_id=26\n- M3 Max 34 records: DB benchmark_record, hardware_id=4\n- Best single: DB record id=1296 (Qwen3.8-9B + IQ2_XS)","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fapple-m4-max-200-tps-club-2026-10-11","2026-10-12T12:12:45.236Z"]