[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f2txk2qsb22ln4":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":36,"bodyZh":40,"bodyEn":41,"sourceName":42,"sourceUrl":43,"updatedAt":44},"open-source-llm-5way-2026-10-09","NEWS","2026-10-09T16:00:00.000Z","开源大模型 5 家横评：Qwen 3.x \u002F DeepSeek V4.1 \u002F GLM-5.3 \u002F Llama 4.5 \u002F Gemma 4","Open-source LLM 5-way comparison: Qwen 3.x \u002F DeepSeek V4.1 \u002F GLM-5.3 \u002F Llama 4.5 \u002F Gemma 4","5 家开源旗舰横评，基于本地 DB RTX 4090 实测：Qwen 3.5 27B 91 tok\u002Fs、Gemma 4 12B 89 tok\u002Fs、Qwen 3.8 Flash 180B 12 tok\u002Fs、DeepSeek V4.1 763B 3 tok\u002Fs。27B 是单卡甜蜜点。","5-way open-source flagship comparison based on local DB RTX 4090 measurements: Qwen 3.5 27B 91 tok\u002Fs, Gemma 4 12B 89 tok\u002Fs, Qwen 3.8 Flash 180B 12 tok\u002Fs, DeepSeek V4.1 763B 3 tok\u002Fs. 27B is the sweet spot.",[12,15,18,21,24,27,30,33],{"id":13,"name":14},7,"DeepSeek-V4-Flash",{"id":16,"name":17},9,"GLM-5.3-Flash",{"id":19,"name":20},11,"Qwen3.8-Flash-Next",{"id":22,"name":23},18,"DeepSeek-V4.1-Flash",{"id":25,"name":26},20,"GLM-5.3",{"id":28,"name":29},50,"Qwen3.5-27B",{"id":31,"name":32},63,"gemma-4-31B-it",{"id":34,"name":35},65,"gemma-4-12B-it",[37],{"id":38,"name":39},1,"NVIDIA RTX 4090","# 开源大模型 5 家横评：Qwen 3.x \u002F DeepSeek V4.1 \u002F GLM-5.3 \u002F Llama 4.5 \u002F Gemma 4\n\n开源阵营现在 5 家——Qwen \u002F DeepSeek \u002F GLM \u002F Llama \u002F Gemma。本帖基于**本地 DB 实测 + 厂商公告**横评。\n\n## 5 家基本盘\n\n| 厂商 | 代表型号 | 参数 | 许可证 |\n|---|---|---|---|\n| Alibaba Qwen | Qwen3.8-Flash-Next | 180B | OPEN_WEIGHTS |\n| Alibaba Qwen | Qwen3.5-27B | 27B | UNVERIFIED |\n| DeepSeek | DeepSeek-V4.1-Flash | 763B | OPEN_SOURCE |\n| DeepSeek | DeepSeek-V4-Flash | 284B | OPEN_SOURCE |\n| Z.ai GLM | GLM-5.3 | 753B | OPEN_WEIGHTS |\n| Z.ai GLM | GLM-5.3-Flash | 320B | OPEN_SOURCE |\n| Meta Llama | Llama 4.5 Maverick | 400B+ | Llama Community |\n| Google Gemma | gemma-4-31B-it | 31B | UNVERIFIED |\n| Google Gemma | gemma-4-12B-it | 12B | UNVERIFIED |\n\n## RTX 4090 实测 decode tok\u002Fs\n\n| 模型 | 参数 | tok\u002Fs |\n|---|---|---|\n| Qwen3.5-27B | 27B | **91.25** |\n| gemma-4-12B-it | 12B | **89.35** |\n| gemma-4-31B-it | 31B | **52.57** |\n| Qwen3.8-Flash-Next | 180B | **11.60** |\n| DeepSeek-V4-Flash | 284B | **5.77** |\n| GLM-5.3-Flash | 320B | **5.25** |\n| GLM-5.3 | 753B | **3.40** |\n| DeepSeek-V4.1-Flash | 763B | **3.30** |\n\n## 5 维评分（1-5 星）\n\n| 维度 | Qwen 3.x | DeepSeek V4 | GLM-5.3 | Llama 4.5 | Gemma 4 |\n|---|---|---|---|---|---|\n| 能力追平闭源 | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ |\n| 本地跑得动（≤30B）| ⭐⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐⭐⭐ |\n| 许可证干净 | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐⭐⭐ |\n| 生态工具支持 | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ |\n| 长期路线图清晰 | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐ | ⭐⭐⭐⭐ |\n\n## 装机推荐\n\n| 你的卡 | 推荐组合 |\n|---|---|\n| 12GB 显存（3060\u002F5070）| Qwen 3.5 27B（91 tok\u002Fs 等效）|\n| 24GB 显存（4090\u002F3090）| Gemma 4 31B（52 tok\u002Fs）|\n| 32GB+ 单卡（5090）| 未来 GLM-5.3 \u002F DeepSeek V4.1 量化版 |\n| 多卡 4×24GB | Qwen3.8-Flash-Next 11.6 tok\u002Fs |\n| 多卡 8×24GB | DeepSeek V4.1 Flash（3.3 tok\u002Fs，长上下文用）|\n\n## 三个判断点\n\n1. 700B+ 模型在单卡几乎不可用——3 tok\u002Fs 体验差\n2. Llama 4.5 在我 DB 还没数据——本帖 Llama 部分基于厂商公告（未实测）\n3. Qwen \"3.7\" 是口语化指代——DB 里实际是 Qwen3.5\u002F3.6\u002F3.8 三个系列\n\n## 证据分层\n\n- 91.25 \u002F 89.35 \u002F 52.57 \u002F 11.60 \u002F 5.77 \u002F 5.25 \u002F 3.40 \u002F 3.30 tok\u002Fs：本地 DB benchmark_record（hardware_id=1 = RTX 4090）\n- 5 维评分：综合 DB 数据 + 厂商公告 + 第三方报告\n- \"Meta 可能不再推下一代\"：Alibaba Apsara 10\u002F2 + 多家复盘（不确定事件）\n\n## 数据来源\n\n- 本地 DB：benchmark_record 表 + model 表\n- Llama 4.5 部分：Meta 官方公告（未实测）","# Open-source LLM 5-way comparison: Qwen 3.x \u002F DeepSeek V4.1 \u002F GLM-5.3 \u002F Llama 4.5 \u002F Gemma 4\n\nFive open-source flagships — Qwen \u002F DeepSeek \u002F GLM \u002F Llama \u002F Gemma. This post is based on **local DB measurements + vendor announcements**.\n\n## Five families\n\n| Vendor | Model | Params | License |\n|---|---|---|---|\n| Alibaba Qwen | Qwen3.8-Flash-Next | 180B | OPEN_WEIGHTS |\n| Alibaba Qwen | Qwen3.5-27B | 27B | UNVERIFIED |\n| DeepSeek | DeepSeek-V4.1-Flash | 763B | OPEN_SOURCE |\n| DeepSeek | DeepSeek-V4-Flash | 284B | OPEN_SOURCE |\n| Z.ai GLM | GLM-5.3 | 753B | OPEN_WEIGHTS |\n| Z.ai GLM | GLM-5.3-Flash | 320B | OPEN_SOURCE |\n| Meta Llama | Llama 4.5 Maverick | 400B+ | Llama Community |\n| Google Gemma | gemma-4-31B-it | 31B | UNVERIFIED |\n| Google Gemma | gemma-4-12B-it | 12B | UNVERIFIED |\n\n## RTX 4090 measured decode tok\u002Fs\n\n| Model | Params | tok\u002Fs |\n|---|---|---|\n| Qwen3.5-27B | 27B | **91.25** |\n| gemma-4-12B-it | 12B | **89.35** |\n| gemma-4-31B-it | 31B | **52.57** |\n| Qwen3.8-Flash-Next | 180B | **11.60** |\n| DeepSeek-V4-Flash | 284B | **5.77** |\n| GLM-5.3-Flash | 320B | **5.25** |\n| GLM-5.3 | 753B | **3.40** |\n| DeepSeek-V4.1-Flash | 763B | **3.30** |\n\n## 5-axis rating (1-5 stars)\n\n| Dimension | Qwen 3.x | DeepSeek V4 | GLM-5.3 | Llama 4.5 | Gemma 4 |\n|---|---|---|---|---|---|\n| Closes-source parity | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ |\n| Local-runnable (≤30B) | ⭐⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐⭐⭐ |\n| License cleanness | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐⭐⭐ |\n| Ecosystem support | ⭐⭐⭐⭐⭐ | ⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐⭐ |\n| Roadmap clarity | ⭐⭐⭐⭐ | ⭐⭐⭐⭐⭐ | ⭐⭐⭐ | ⭐⭐ | ⭐⭐⭐⭐ |\n\n## Build recommendations\n\n| Your card | Pick |\n|---|---|\n| 12GB (3060\u002F5070) | Qwen 3.5 27B (91 tok\u002Fs equivalent) |\n| 24GB (4090\u002F3090) | Gemma 4 31B (52 tok\u002Fs) |\n| 32GB+ single (5090) | Future GLM-5.3 \u002F DeepSeek V4.1 quantized |\n| Multi 4×24GB | Qwen3.8-Flash-Next 11.6 tok\u002Fs |\n| Multi 8×24GB | DeepSeek V4.1 Flash (3.3 tok\u002Fs, long context) |\n\n## Three judgment points\n\n1. 700B+ models nearly unusable on a single card — 3 tok\u002Fs is poor\n2. Llama 4.5 has no local DB data — Llama section based on vendor announcements (not measured)\n3. \"Qwen 3.7\" is colloquial — actual DB models are Qwen 3.5\u002F3.6\u002F3.8 series\n\n## Evidence\n\n- 91.25 \u002F 89.35 \u002F 52.57 \u002F 11.60 \u002F 5.77 \u002F 5.25 \u002F 3.40 \u002F 3.30 tok\u002Fs: local DB benchmark_record (hardware_id=1 = RTX 4090)\n- 5-axis rating: combined DB data + vendor announcements + third-party reports\n- \"Meta may not ship next gen\": Alibaba Apsara 10\u002F2 + multiple recaps (uncertain event)\n\n## Sources\n\n- Local DB: benchmark_record table + model table\n- Llama 4.5: Meta official announcement (not measured)","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fopen-source-llm-5way-2026-10-09","2026-10-10T05:41:10.815Z"]