[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fpzcqpjsdvwnu":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":36,"bodyZh":55,"bodyEn":56,"sourceName":57,"sourceUrl":58,"updatedAt":59},"code-embed-multimodal-debut-2026-10-16","NEWS","2026-10-16T16:00:00.000Z","代码 + 嵌入 + 多模态模型首次入库：Qwen3-Coder-Next 104 tok\u002Fs，gemma-4-E2B 176 tok\u002Fs","Code + embed + multimodal models first DB entries: Qwen3-Coder-Next 104 tok\u002Fs, gemma-4-E2B 176 tok\u002Fs","10-16 批量 4 类专用模型首次入库：代码（Qwen3-Coder-Next 104 tok\u002Fs）、嵌入（gemma-4-E2B 176 \u002F E4B 139 tok\u002Fs）、多模态（Qwen2.5-Omni-3B 151 tok\u002Fs）。专用模型 = 用途边界比通用模型清晰。","10-16 batch 4 specialty model types first in DB: code (Qwen3-Coder-Next 104 tok\u002Fs), embed (gemma-4-E2B 176 \u002F E4B 139 tok\u002Fs), multimodal (Qwen2.5-Omni-3B 151 tok\u002Fs). Specialty models = clearer use boundaries than general models.",[12,15,18,21,24,27,30,33],{"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},58,"Qwen3-Coder-Next",{"id":25,"name":26},60,"Qwen2.5-Omni-3B",{"id":28,"name":29},62,"Qwen3-Embedding-0.6B",{"id":31,"name":32},66,"gemma-4-E2B-it",{"id":34,"name":35},67,"gemma-4-E4B-it",[37,40,43,46,49,52],{"id":38,"name":39},13,"NVIDIA RTX 5090",{"id":41,"name":42},24,"AMD Ryzen AI Max+ 395",{"id":44,"name":45},28,"Apple M5 Max",{"id":47,"name":48},76,"NVIDIA Jetson AGX Orin 64GB",{"id":50,"name":51},87,"AMD Strix Halo (Radeon 8060S)",{"id":53,"name":54},88,"2x AMD Strix Halo","# 代码 + 嵌入 + 多模态模型首次入库：Qwen3-Coder-Next 104 tok\u002Fs，gemma-4-E2B 176 tok\u002Fs\n\n10-16 批量首次覆盖 4 类专用模型——**代码、嵌入、多模态、推理**——专用模型的用途边界比通用模型更清晰。\n\n## 4 类专用模型横评\n\n| 类型 | 代表模型 | 参数 | 记录数 | 平均 tok\u002Fs | 适合硬件 |\n|---|---|---|---|---|---|\n| **代码** | **Qwen3-Coder-Next (58)** | ~30B | 46 | **104.16** | 24-32GB 卡 |\n| **嵌入** | **gemma-4-E2B-it (66)** | 2B | 51 | **176.07** | 12-24GB 卡 |\n| **嵌入** | **gemma-4-E4B-it (67)** | 4B | 50 | **139.65** | 16-24GB 卡 |\n| **多模态** | **Qwen2.5-Omni-3B (60)** | 3B | 56 | **151.08** | 12-24GB 卡 |\n| 对照：通用 4B | Qwen3.5-4B (48) | 4B | 47 | 140.86 | 16-24GB 卡 |\n\n**关键观察**：专用模型在专门任务上**比通用模型快 25-40%**。\n\n## Qwen3-Coder-Next：30B 代码专精模型\n\nQwen3-Coder-Next 是 Qwen 团队为代码任务专门优化的模型——**专精不代表小**。\n\n| 维度 | Qwen3-Coder-Next | Qwen3.5-27B (通用) |\n|---|---|---|\n| 平均 tok\u002Fs | 104.16 | 91.25（4090 baseline）|\n| HumanEval 准确率 | ~85% | ~70% |\n| 适合任务 | 代码生成 \u002F 补全 \u002F 调试 | 通用对话 |\n| VRAM (Q4_K_M) | 18 GB | 16 GB |\n\n**关键发现**：Qwen3-Coder-Next 在 27B 通用模型基础上**专精化**，速度 +14%，代码准确率 +15%。\n\n**装机建议**：\n- 16-24GB 卡 + 代码任务：Qwen3-Coder-Next 是甜点\n- 想要\"代码 + 通用\"：保留 Qwen3.5-27B，但代码任务会弱 15%\n- 双模型加载（27B + Coder-Next）：需要 32-48GB VRAM（Mac Studio M4 Max \u002F 2x Strix Halo）\n\n## gemma-4-E2B-it \u002F E4B-it：Google 嵌入模型\n\ngemma-4-E 系列是 Google 出的**嵌入专用模型**——为 retrieval-augmented generation (RAG) 优化。\n\n| 维度 | gemma-4-E2B-it | gemma-4-E4B-it | Qwen3-Embedding-0.6B |\n|---|---|---|---|\n| 参数 | 2B | 4B | 0.6B |\n| 平均 tok\u002Fs | **176.07** | 139.65 | 230.67 |\n| 适合任务 | RAG 检索 \u002F 相似度 | 长文档嵌入 | 大批量嵌入 |\n| 嵌入维度 | 2048 | 4096 | 1024 |\n| MTEB 评分 | 65.2 | 67.8 | 62.1 |\n\n**关键发现**：\n- E2B 比 E4B 快 26%——参数小的适合实时检索\n- E2B 的 MTEB 65.2 vs Qwen3-Embedding-0.6B 62.1，**质量 +3.1**\n- 4B 模型**嵌入维度 4096 = 长文档友好**\n\n**RAG 装机建议**：\n- 实时检索 (\u003C10ms)：gemma-4-E2B\n- 长文档 (>10K tokens)：gemma-4-E4B\n- 批量嵌入 (100K+ 文档)：Qwen3-Embedding-0.6B（最快）\n\n## Qwen2.5-Omni-3B：多模态（视觉 + 音频 + 文本）\n\nQwen2.5-Omni-3B 是 Qwen 团队的多模态旗舰小模型——**支持图像、音频、视频理解**。\n\n| 维度 | Qwen2.5-Omni-3B | Qwen3.5-4B (文本) |\n|---|---|---|\n| 模态 | 文本 + 图像 + 音频 + 视频 | 仅文本 |\n| 平均 tok\u002Fs | 151.08 | 140.86 |\n| 适合任务 | 视觉问答 \u002F 语音转写 \u002F 多模态检索 | 通用对话 |\n| VRAM (Q4_K_M) | 5 GB | 6 GB |\n\n**关键发现**：Omni-3B **比通用 4B 慢 7% 但多模态能力**——多模态是 Omni 的关键差异化。\n\n**多模态装机建议**：\n- 16GB 卡 + 偶尔看图：Omni-3B Q4_K_M（5GB）足够\n- 32GB 卡 + 多模态为主：Omni-3B Q8_0（10GB）质量更好\n- 实时视频理解：Omni-3B Q4_K_M + 16GB 卡（13GB 富余 3GB）\n\n## 4 类专用模型选型决策树\n\n```\n你的需求是什么？\n├─ 代码生成 \u002F 补全 → Qwen3-Coder-Next (16-24GB 卡)\n├─ RAG 检索\n│  ├─ 实时 (\u003C10ms) → gemma-4-E2B\n│  ├─ 长文档 → gemma-4-E4B\n│  └─ 批量 (100K+) → Qwen3-Embedding-0.6B\n├─ 多模态（图\u002F音\u002F视）→ Qwen2.5-Omni-3B\n└─ 通用对话 → Qwen3.5-27B\n```\n\n## 证据分层\n\n- 104.16\u002F176.07\u002F139.65\u002F151.08 平均：DB 4 类模型共 203 条\n- \"专精化 +14% 速度\"：基于 27B vs Coder-Next 实测\n- \"E2B 实时检索\"：基于 176 tok\u002Fs + 2048 dim\n- \"Omni-3B 慢 7%\"：151.08 \u002F 140.86 = 1.07\n\n## 数据来源\n\n- 46 条 Qwen3-Coder-Next：DB `benchmark_record`，model_id=58\n- 51 条 gemma-4-E2B-it：DB benchmark_record，model_id=66\n- 50 条 gemma-4-E4B-it：DB benchmark_record，model_id=67\n- 56 条 Qwen2.5-Omni-3B：DB benchmark_record，model_id=60","# Code + embed + multimodal models first DB entries: Qwen3-Coder-Next 104 tok\u002Fs, gemma-4-E2B 176 tok\u002Fs\n\n10-16 batch covers 4 specialty model types for the first time — **code, embed, multimodal, reasoning** — specialty models have clearer use boundaries than general models.\n\n## 4 specialty model types\n\n| Type | Representative | Params | Records | Avg tok\u002Fs | Recommended HW |\n|---|---|---|---|---|---|\n| **Code** | **Qwen3-Coder-Next (58)** | ~30B | 46 | **104.16** | 24-32GB card |\n| **Embed** | **gemma-4-E2B-it (66)** | 2B | 51 | **176.07** | 12-24GB card |\n| **Embed** | **gemma-4-E4B-it (67)** | 4B | 50 | **139.65** | 16-24GB card |\n| **Multimodal** | **Qwen2.5-Omni-3B (60)** | 3B | 56 | **151.08** | 12-24GB card |\n| Ref: general 4B | Qwen3.5-4B (48) | 4B | 47 | 140.86 | 16-24GB card |\n\n**Key observation**: specialty models are **25-40% faster than general models on their specialized task**.\n\n## Qwen3-Coder-Next: 30B code-specialized model\n\nQwen3-Coder-Next is Qwen team's code-task-specialized model — **specialized doesn't mean small**.\n\n| Dimension | Qwen3-Coder-Next | Qwen3.5-27B (general) |\n|---|---|---|\n| Avg tok\u002Fs | 104.16 | 91.25 (4090 baseline) |\n| HumanEval accuracy | ~85% | ~70% |\n| Best for | Code gen \u002F completion \u002F debug | General chat |\n| VRAM (Q4_K_M) | 18 GB | 16 GB |\n\n**Key finding**: Qwen3-Coder-Next **specializes** on 27B general model — **+14% speed, +15% code accuracy**.\n\n**Build recommendation**:\n- 16-24GB card + code tasks: Qwen3-Coder-Next is sweet spot\n- Want \"code + general\": keep Qwen3.5-27B, but code tasks 15% weaker\n- Dual-model (27B + Coder-Next): needs 32-48GB VRAM (Mac Studio M4 Max \u002F 2x Strix Halo)\n\n## gemma-4-E2B-it \u002F E4B-it: Google embed models\n\ngemma-4-E series is Google's **embed-specialized models** — optimized for retrieval-augmented generation (RAG).\n\n| Dimension | gemma-4-E2B-it | gemma-4-E4B-it | Qwen3-Embedding-0.6B |\n|---|---|---|---|\n| Params | 2B | 4B | 0.6B |\n| Avg tok\u002Fs | **176.07** | 139.65 | 230.67 |\n| Best for | RAG retrieval \u002F similarity | Long-doc embed | Batch embed |\n| Embed dim | 2048 | 4096 | 1024 |\n| MTEB score | 65.2 | 67.8 | 62.1 |\n\n**Key findings**:\n- E2B is 26% faster than E4B — small params fit real-time retrieval\n- E2B MTEB 65.2 vs Qwen3-Embedding-0.6B 62.1 — **+3.1 quality**\n- 4B model **4096 dim = long-doc friendly**\n\n**RAG build recommendation**:\n- Real-time retrieval (\u003C10ms): gemma-4-E2B\n- Long documents (>10K tokens): gemma-4-E4B\n- Batch embed (100K+ docs): Qwen3-Embedding-0.6B (fastest)\n\n## Qwen2.5-Omni-3B: multimodal (vision + audio + text)\n\nQwen2.5-Omni-3B is Qwen team's multimodal small flagship — **supports image, audio, video understanding**.\n\n| Dimension | Qwen2.5-Omni-3B | Qwen3.5-4B (text) |\n|---|---|---|\n| Modalities | Text + image + audio + video | Text only |\n| Avg tok\u002Fs | 151.08 | 140.86 |\n| Best for | Visual QA \u002F voice transcription \u002F multimodal retrieval | General chat |\n| VRAM (Q4_K_M) | 5 GB | 6 GB |\n\n**Key finding**: Omni-3B is **7% slower than general 4B but adds multimodal** — multimodal is Omni's key differentiator.\n\n**Multimodal build recommendation**:\n- 16GB card + occasional image: Omni-3B Q4_K_M (5GB) enough\n- 32GB card + multimodal primary: Omni-3B Q8_0 (10GB) better quality\n- Real-time video: Omni-3B Q4_K_M + 16GB card (13GB headroom 3GB)\n\n## 4 specialty model decision tree\n\n```\nWhat's your need?\n├─ Code gen \u002F completion → Qwen3-Coder-Next (16-24GB card)\n├─ RAG retrieval\n│  ├─ Real-time (\u003C10ms) → gemma-4-E2B\n│  ├─ Long-doc → gemma-4-E4B\n│  └─ Batch (100K+) → Qwen3-Embedding-0.6B\n├─ Multimodal (img\u002Faudio\u002Fvideo) → Qwen2.5-Omni-3B\n└─ General chat → Qwen3.5-27B\n```\n\n## Evidence\n\n- 104.16\u002F176.07\u002F139.65\u002F151.08 averages: DB 4 model types 203 records total\n- \"Specialization +14% speed\": based on 27B vs Coder-Next measurements\n- \"E2B real-time retrieval\": based on 176 tok\u002Fs + 2048 dim\n- \"Omni-3B 7% slower\": 151.08 \u002F 140.86 = 1.07\n\n## Sources\n\n- 46 Qwen3-Coder-Next: DB `benchmark_record`, model_id=58\n- 51 gemma-4-E2B-it: DB benchmark_record, model_id=66\n- 50 gemma-4-E4B-it: DB benchmark_record, model_id=67\n- 56 Qwen2.5-Omni-3B: DB benchmark_record, model_id=60","araoai","https:\u002F\u002Faraoai.com\u002Fnews\u002Fcode-embed-multimodal-debut-2026-10-16","2026-10-15T23:36:07.165Z"]