[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1sswpq2c2w0ij":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":18,"bodyZh":22,"bodyEn":23,"sourceName":24,"sourceUrl":25,"updatedAt":26},"open-close-gap-2026-10","NEWS","2026-10-06T10:00:00.000Z","4 月开源闭源差距「个位数百分点」· 10 月 Gemini 4 Argon 又超了","Open vs closed-source gap closed to single digits in April; Gemini 4 Argon retook the lead in October","4 月开源 vs 闭源在主流 benchmark 差距缩到 \u003C10%。10 月 Gemini 4 Argon \u002F Claude Opus 5.5 重新超越；但 80% 任务开源旗舰够用。","Open-weight vs closed-source gap narrowed to single digits in April. Gemini 4 Argon and Claude Opus 5.5 retook the lead in October, but 80% of tasks are still served well by top open models.",[12,15],{"id":13,"name":14},18,"DeepSeek-V4.1-Flash",{"id":16,"name":17},50,"Qwen3.5-27B",[19],{"id":20,"name":21},1,"NVIDIA RTX 4090","# 4 月开源闭源差距「个位数百分点」· 10 月 Gemini 4 Argon 又超了\n\n这事很多人没看懂——不是说开源赢了，而是说**对 80% 的任务，开源已经够了**。\n\n## 10 月开源 vs 闭源\n\n| 模型 | 厂商 | 类型 | 当前状态 |\n|------|------|------|----------|\n| Gemini 4 Argon | Google | 闭源 | 10\u002F1 发布 · #1 Text Arena |\n| Claude Sonnet 5.5 | Anthropic | 闭源 | 10\u002F2 · 30% 快 30% 便宜 |\n| GPT-6.1 Sol | OpenAI | 闭源 | 10\u002F4 · $2\u002F$10 |\n| Claude Opus 5.5 | Anthropic | 闭源 | 4 月 |\n| Qwen 3.7 | Alibaba | Apache 2.0 | H1 旗舰；Qwen 4 训练中 |\n| DeepSeek V4.1 Flash | DeepSeek | MIT | 已发 · 1.6T \u002F 49B active |\n| GLM-5.3 | Z.ai | MIT | Anthropic 称\"最 cyber-capable\" |\n| Llama 4.5 Maverick | Meta | Llama Community | Meta 可能**不再推**下一代 |\n| Gemma 4 31B | Google | Apache 2.0 | 干净许可证 |\n\n## 4 月发生了什么\n\nThorsten Meyer 4 月那篇《Single Digits》：开源 vs 闭源在主流 benchmark 上差距从 20-30% 缩到「个位数百分点」（\u003C 10%）。\n\n关键不是开源\"赢了\"——关键是**能力追平之后，开源多了三件事**：\n\n- ✅ 私有数据不出门（最关键）\n- ✅ 调用成本边际 0\n- ✅ 可微调 \u002F 可离线 \u002F 可审计\n\n10 月 Gemini 4 Argon 重新超越闭源 leaderboard——但对 80% 任务（写代码 \u002F 翻译 \u002F 总结 \u002F 抽取）开源旗舰够用。\n\n## 经济账（写代码场景 200 万 tokens\u002F月）\n\n闭源路径：\n- Gemini 4 Argon：$2\u002F$10\u002FM × 2M = $20\u002F月 = ¥144\u002F月\n- 一年：¥1728\n\n本地路径（Qwen 3.5 27B Q4_K_M）：\n- 一次性硬件：¥9000（4090 多卡 + 部分 offload）\n- 月电费：¥150\n- 一年回本 vs Gemini API：约 4 年\n- 之后边际 0\n\n## 三个判断点\n\n1️⃣ 闭源 leaderboard 拉锯≠你用得上。Gemini 4 Argon 的 1M 输出你 90% 用不到\n2️⃣ 开源 ≠ 免费。Qwen 3.5 27B ≈ 130GB 装下，4090 单卡跑要分块，**真全本地要 64GB+ 卡**\n3️⃣ Meta 不推下一代这事——开源阵营少一个大型玩家。Qwen \u002F DeepSeek \u002F GLM 顶上\n\n## 证据分层\n\n- \"差距个位数百分点\"：4 月 thorstenmeyerai 复盘 H1 数据\n- \"Gemini 4 Argon #1\"：10\u002F1 Google 官方 + AutomationBench \u002F LVBench 第三方\n- \"GLM-5.3 cyber-capable\"：Anthropic 9 月报告措辞\n- \"Qwen 4 训练中\"：Alibaba Apsara 10\u002F2 大会\n- \"Meta 不推下一代\"：10\u002F2 Apsara + 多家复盘交叉验证","# Open vs closed-source gap closed to single digits in April; Gemini 4 Argon retook the lead in October\n\nThe story is not \"open-source won\" — it's that **for 80% of tasks, open weights are already enough**.\n\n## October state\n\n| Model | Vendor | Type | Status |\n|------|------|------|----------|\n| Gemini 4 Argon | Google | Closed | 10\u002F1 launch · #1 Text Arena |\n| Claude Sonnet 5.5 | Anthropic | Closed | 10\u002F2 · 30% faster, 30% cheaper |\n| GPT-6.1 Sol | OpenAI | Closed | 10\u002F4 · $2\u002F$10 |\n| Claude Opus 5.5 | Anthropic | Closed | April |\n| Qwen 3.7 | Alibaba | Apache 2.0 | H1 flagship; Qwen 4 in training |\n| DeepSeek V4.1 Flash | DeepSeek | MIT | Released · 1.6T \u002F 49B active |\n| GLM-5.3 | Z.ai | MIT | Anthropic called \"most cyber-capable\" |\n| Llama 4.5 Maverick | Meta | Llama Community | Meta may **not ship** a next gen |\n| Gemma 4 31B | Google | Apache 2.0 | Clean license |\n\n## What happened in April\n\nThorsten Meyer's April piece \"Single Digits\": open vs closed on mainstream benchmarks narrowed from 20-30% to single digits (\u003C10%).\n\nThe point is not \"open won\" — it's that **once capability catches up, open weights give you three more things**:\n\n- ✅ Private data stays in house\n- ✅ Marginal cost of calls → 0\n- ✅ Tunable \u002F offline \u002F auditable\n\nGemini 4 Argon retook the closed-source lead in October — but for 80% of tasks (coding \u002F translation \u002F summarization \u002F extraction) the top open-weight flagships are enough.\n\n## Three judgment points\n\n1️⃣ Closed-source leaderboard churn ≠ what you actually use. You'll use 10% of Gemini 4 Argon's 1M output\n2️⃣ Open ≠ free. Qwen 3.5 27B is ~130GB, single 4090 needs sharding, **real full-local needs 64GB+ cards**\n3️⃣ Meta not shipping next gen — open camp loses a major player. Qwen \u002F DeepSeek \u002F GLM fill in","thorstenmeyerai.com","https:\u002F\u002Fthorstenmeyerai.com\u002Finsights\u002Fsingle-digits-the-april-that-closed-the-open-weight-gap","2026-10-09T06:21:02.105Z"]