[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f12nhwgwqx2nmo":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":48,"bodyZh":70,"bodyEn":71,"sourceName":72,"sourceUrl":73,"updatedAt":74},"apple-m1-m2-revival-llm-2026-10-16","NEWS","2026-10-16T10:00:00.000Z","Apple M1\u002FM2 复活：旧 Mac 也能跑 LLM——M1 平均 90 tok\u002Fs，M2 Ultra 192GB 跑 70B","Apple M1\u002FM2 revival: old Macs can run LLM — M1 avg 90 tok\u002Fs, M2 Ultra 192GB runs 70B","10-16 批量 209 条 Apple M1-M2 实测首次入库：M1 90.65、M1 Pro 118.71、M1 Max 147.03、M2 Pro 123.14、M2 Ultra 172.06 tok\u002Fs。**M2 Ultra 192GB unified 是 ¥24000 价位唯一能跑 70B 的整机**。","10-16 batch 209 Apple M1-M2 measurements first time in DB: M1 90.65, M1 Pro 118.71, M1 Max 147.03, M2 Pro 123.14, M2 Ultra 172.06 tok\u002Fs. **M2 Ultra 192GB unified is the only ¥24000 whole-machine that runs 70B**.",[12,15,18,21,24,27,30,33,36,39,42,45],{"id":13,"name":14},48,"Qwen3.5-4B",{"id":16,"name":17},49,"Qwen3.5-9B",{"id":19,"name":20},50,"Qwen3.5-27B",{"id":22,"name":23},51,"Qwen3.5-0.8B",{"id":25,"name":26},52,"Qwen3.8-9B",{"id":28,"name":29},53,"Qwen3-8B",{"id":31,"name":32},54,"Qwen3-4B",{"id":34,"name":35},55,"Qwen3-1.7B",{"id":37,"name":38},58,"Qwen3-Coder-Next",{"id":40,"name":41},60,"Qwen2.5-Omni-3B",{"id":43,"name":44},65,"gemma-4-12B-it",{"id":46,"name":47},66,"gemma-4-E2B-it",[49,52,55,58,61,64,67],{"id":50,"name":51},4,"Apple M3 Max",{"id":53,"name":54},28,"Apple M5 Max",{"id":56,"name":57},99,"Apple M1",{"id":59,"name":60},100,"Apple M1 Pro",{"id":62,"name":63},101,"Apple M1 Max",{"id":65,"name":66},102,"Apple M2 Pro",{"id":68,"name":69},103,"Apple M2 Ultra","# Apple M1\u002FM2 复活：旧 Mac 也能跑 LLM——M1 平均 90 tok\u002Fs，M2 Ultra 192GB 跑 70B\n\n10-16 批量入库 209 条 Apple M1\u002FM2 系列 L2 实测——**Apple 2020-2023 旧 Mac 生态首次进入 DB**。\n\n## 整体复盘（5 个 M 系芯片）\n\n| 硬件 | VRAM | 记录数 | 平均 tok\u002Fs | 峰值 | 形态 |\n|---|---|---|---|---|---|\n| **M2 Ultra** | **192GB** | 47 | **172.06** | 393.85 | Mac Studio \u002F Mac Pro |\n| M1 Max | 64GB | 41 | 147.03 | **395.33** | MBP 16 \u002F Mac Studio |\n| M2 Pro | 32GB | 41 | 123.14 | 271.73 | MBP 14 \u002F Mac mini |\n| M1 Pro | 32GB | 41 | 118.71 | 263.91 | MBP 14\u002F16 |\n| M1 | 16GB | 39 | 90.65 | 231.36 | MBP 13 \u002F Mac mini \u002F iPad |\n| 同比 M5 Max (34 条) | 128GB | — | 257.67 | 366.43 | Mac Studio |\n| 同比 M4 Max (47 条) | 128GB | — | 196.99 | 353.90 | Mac Studio |\n\n**关键观察**：\n1. **M1 都能跑 90 tok\u002Fs**——2020 年的 MBP 13 + M1 + 16GB unified 跑 Qwen3.5-27B IQ2_XS，足够日常使用\n2. **M2 Ultra 192GB unified** 是 ¥24000 价位**唯一能跑 70B 量化版**的整机（vs Mac Studio M5 Max 128GB ¥24000 + DGX Spark 128GB ¥30000）\n3. **M1 Max 峰值 395.33 tok\u002Fs**——Apple Silicon 老旗舰仍然站得住 200+ tok\u002Fs 俱乐部\n\n## M2 Ultra 192GB 杀手锏\n\n| 模型 | 量化 | VRAM 占用 | tok\u002Fs |\n|---|---|---|---|\n| Qwen3.5-27B | Q4_K_M | 16 GB | 130 |\n| Qwen3.5-27B | IQ1_S | 13 GB | 175 |\n| **DeepSeek V4.1 Flash 763B** | **IQ1_S** | **150+ GB** | **3-4** |\n| **GLM-5.3 753B** | **IQ2_XS** | **130+ GB** | **5-7** |\n| Qwen3.5-27B + gemma-4-31B | 双加载 | 32 GB | 130 \u002F 80 |\n\n**M2 Ultra 192GB 的独特点**：能装下 V4.1 Flash 763B IQ1_S（150GB+） + 同时给 Mac 跑 macOS Sonoma，**还有 40GB+ 富余**。M5 Max 128GB 装不下 700B+。\n\n## 旧 Mac 复活清单\n\n| 你的旧 Mac | 适合度 | 推荐模型 | tok\u002Fs |\n|---|---|---|---|\n| MBP 13 M1 16GB | ✅ 强 | Qwen3.5-4B IQ4_XS | 90 |\n| MBP 14 M1 Pro 32GB | ✅ 强 | Qwen3.5-9B Q4_K_M | 118 |\n| MBP 16 M1 Max 64GB | ✅ 很强 | Qwen3.5-27B IQ2_XS | 147 |\n| Mac mini M2 Pro 32GB | ✅ 强 | Qwen3.5-9B Q4_K_M | 123 |\n| Mac Studio M2 Ultra 192GB | ⭐ 顶级 | DeepSeek V4.1 Flash 763B | 3.5 |\n\n**装机建议**：**别换 Mac，4 年前的 M1 Max 64GB + Qwen3.5-27B = 147 tok\u002Fs 足够本地推理**。M1-M2 不是电子垃圾，是 LLM 推理的甜区。\n\n## 证据分层\n\n- 90.65\u002F118.71\u002F147.03\u002F123.14\u002F172.06 平均：DB 209 条\n- M1 Max 峰值 395.33：DB record id=2044\n- M2 Ultra 192GB：Apple 官方 spec\n- \"M1 都能跑 90\"：DB 39 条数据\n- \"M2 Ultra 唯一 ¥24000 跑 70B\"：基于当前价格 + VRAM 对比\n\n## 数据来源\n\n- 209 条 M1-M2 记录：DB `benchmark_record` 表，hardware_id IN (99, 100, 101, 102, 103)\n- M5 Max 34 条：DB benchmark_record，hardware_id=28\n- M4 Max 47 条：DB benchmark_record，hardware_id=4","# Apple M1\u002FM2 revival: old Macs can run LLM — M1 avg 90 tok\u002Fs, M2 Ultra 192GB runs 70B\n\n10-16 batch has 209 Apple M1\u002FM2-series L2 records — **Apple 2020-2023 old Mac ecosystem first time in DB**.\n\n## Overall recap (5 M-series chips)\n\n| Hardware | VRAM | Records | Avg tok\u002Fs | Peak | Form factor |\n|---|---|---|---|---|---|\n| **M2 Ultra** | **192GB** | 47 | **172.06** | 393.85 | Mac Studio \u002F Mac Pro |\n| M1 Max | 64GB | 41 | 147.03 | **395.33** | MBP 16 \u002F Mac Studio |\n| M2 Pro | 32GB | 41 | 123.14 | 271.73 | MBP 14 \u002F Mac mini |\n| M1 Pro | 32GB | 41 | 118.71 | 263.91 | MBP 14\u002F16 |\n| M1 | 16GB | 39 | 90.65 | 231.36 | MBP 13 \u002F Mac mini \u002F iPad |\n| For ref: M5 Max (34 records) | 128GB | — | 257.67 | 366.43 | Mac Studio |\n| For ref: M4 Max (47 records) | 128GB | — | 196.99 | 353.90 | Mac Studio |\n\n**Key observations**:\n1. **M1 hits 90 tok\u002Fs** — 2020 MBP 13 + M1 + 16GB unified runs Qwen3.5-27B IQ2_XS, enough for daily use\n2. **M2 Ultra 192GB unified** is the **only ¥24000 whole-machine that runs 70B quantized** (vs Mac Studio M5 Max 128GB ¥24000 + DGX Spark 128GB ¥30000)\n3. **M1 Max peak 395.33 tok\u002Fs** — Apple Silicon old flagship still in 200+ tok\u002Fs club\n\n## M2 Ultra 192GB killer feature\n\n| Model | Quant | VRAM used | tok\u002Fs |\n|---|---|---|---|\n| Qwen3.5-27B | Q4_K_M | 16 GB | 130 |\n| Qwen3.5-27B | IQ1_S | 13 GB | 175 |\n| **DeepSeek V4.1 Flash 763B** | **IQ1_S** | **150+ GB** | **3-4** |\n| **GLM-5.3 753B** | **IQ2_XS** | **130+ GB** | **5-7** |\n| Qwen3.5-27B + gemma-4-31B | Dual-load | 32 GB | 130 \u002F 80 |\n\n**M2 Ultra 192GB unique point**: fits V4.1 Flash 763B IQ1_S (150GB+) + run macOS Sonoma simultaneously, **with 40GB+ headroom**. M5 Max 128GB can't fit 700B+.\n\n## Old Mac revival checklist\n\n| Your old Mac | Fit | Recommended model | tok\u002Fs |\n|---|---|---|---|\n| MBP 13 M1 16GB | ✅ strong | Qwen3.5-4B IQ4_XS | 90 |\n| MBP 14 M1 Pro 32GB | ✅ strong | Qwen3.5-9B Q4_K_M | 118 |\n| MBP 16 M1 Max 64GB | ✅ very strong | Qwen3.5-27B IQ2_XS | 147 |\n| Mac mini M2 Pro 32GB | ✅ strong | Qwen3.5-9B Q4_K_M | 123 |\n| Mac Studio M2 Ultra 192GB | ⭐ top | DeepSeek V4.1 Flash 763B | 3.5 |\n\n**Build recommendation**: **don't upgrade the Mac, 4-year-old M1 Max 64GB + Qwen3.5-27B = 147 tok\u002Fs is enough for local inference**. M1-M2 isn't e-waste, it's the LLM inference sweet spot.\n\n## Evidence\n\n- 90.65\u002F118.71\u002F147.03\u002F123.14\u002F172.06 averages: DB 209 records\n- M1 Max peak 395.33: DB record id=2044\n- M2 Ultra 192GB: Apple official spec\n- \"M1 hits 90\": DB 39 records\n- \"M2 Ultra is the only ¥24000 70B\": based on current pricing + VRAM comparison\n\n## Sources\n\n- 209 M1-M2 records: DB `benchmark_record` table, hardware_id IN (99, 100, 101, 102, 103)\n- M5 Max 34 records: DB benchmark_record, hardware_id=28\n- M4 Max 47 records: DB benchmark_record, hardware_id=4","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fapple-m1-m2-revival-llm-2026-10-16","2026-10-15T23:36:07.147Z"]