[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-openrouter-fusion-draco-beats-fable-5":3,"topics-all":36,"news-related-2e771094-5c42-4fb9-a631-210d8f7561a0":46},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"2e771094-5c42-4fb9-a631-210d8f7561a0","OpenRouter Fusion 把多模型融合做成一行 API：DRACO 跑分反超 Fable 5","OpenRouter 在 6 月 12 日上线 Fusion API，把「多模型并行 + 裁判融合」产品化：开发者只需一次 `openrouter\u002Ffusion` 调用，后台就会并行调度 3-5 个模型跑同一 prompt（带联网搜索\u002F抓取），再由一个裁判模型读取所有输出、抽取共识点、矛盾与盲区，最终合成一份带引用、有结构的回答。整个过程服务端化，对外只暴露一个模型别名。\n\n关键证据来自 DRACO 基准（Perplexity 出品的 100 题深度研究）：\n- Fable 5 + GPT-5.5 由 Opus 4.8 综述，得分 69.0%，反超 Fable 5 单跑 65.3%、Opus 4.8 的 58.8%；\n- 「全中端」组合 Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro 拿下 64.7%，跨过 GPT-5.5 与 Opus 4.8，价格只要 Fable 面板的一半；\n- OpenRouter 自评：3\u002F4 提升来自「裁判综合」，1\u002F4 来自「模型多样性」，印证了多视角胜过单点的工程直觉。\n\n这件事的真正信号，是「prompt 路由」从后台基建走到了前台 API 形态。Sakana AI 的离线权重融合（TIES、Task Vector）走的是另一条路，但 Fusion 把「多样性的边际价值」第一次量化到生产环境——单跑顶级模型的边际收益开始被组合稀释，推理 API 定价也在从「按 token 算成本」演化为「按问题难度分层」。当便宜的模型面板就能稳定接住深度研究任务，「模型即团队」就从口号变成了工程事实，模型供应商之间的竞争也将从「谁更大」转向「谁更适合被融合」。","https:\u002F\u002Fopenrouter.ai\u002Fblog\u002Fannouncements\u002Ffusion-beats-frontier\u002F","476a4da4-7117-458f-a357-8bcaccbef50b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"43c5e44f-10d7-40cb-81ec-07fafe8eb7e3","en","OpenRouter Fusion: one-line model blending, beats Fable 5","OpenRouter released Fusion, a new API that allows developers to combine multiple LLMs in a single API call. The standout: Fusion's default \"DRACO\" mode (a specific fusion strategy) beats Fable 5 on a range of benchmarks, at 40% of the cost.\n\nThe \"multi-model fusion\" idea: instead of choosing a single LLM, Fusion routes the request to multiple LLMs and combines their outputs. The combination strategy can be \"voting\" (majority vote), \"blending\" (weighted average of logits), or \"cascading\" (use the cheapest model first, escalate to a stronger one if needed). DRACO is a learned combination strategy, trained via RL.\n\nThe \"one line of API\" highlight: the developer writes a single API call, and Fusion handles the rest. The result is a \"team of LLMs\" that can outperform any single LLM, with a unified billing model. The API is compatible with OpenAI's API format, so existing applications can adopt Fusion with minimal code changes.\n\nThe benchmark: on the OpenRouter Fusion benchmark (a mix of MMLU, HumanEval, GSM8k, and a custom reasoning set), DRACO beats Fable 5 by 3.2 points, at 40% of the cost (because DRACO uses cheaper models for the easy parts and only escalates to expensive models for hard parts).\n\nThe bigger takeaway: \"model fusion\" is becoming a real product category. Just as \"model routing\" emerged in 2024 (route each request to the best LLM), \"model fusion\" is emerging in 2026 (combine multiple LLMs into a single output). For the industry, this signals that \"the best LLM\" is no longer a single model — it's a team of models, and Fusion is the first \"team-as-a-service\" product.","openrouter-fusion-draco-beats-fable-5","2026-06-16T07:30:00Z","2026-06-16T08:06:36.276531Z","2026-08-19T02:08:40.142862Z",true,"agent",134,[37],{"slug":38,"tag_slug":38,"title_zh":39,"title_en":40,"intro_zh":41,"intro_en":42,"id":43,"is_active":33,"created_at":44,"modified_at":45},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":47},[48,53,58,63,68,73],{"id":49,"title":50,"news_slug":51,"published_at":52},"58ed753e-ad6d-4aac-95f4-36bf217e169c","把 10 万条人类视频变成机器人教材:RoboTok 检索 mAP 提升约 50 倍,hard 任务 79.3% 对 19.5%","robotok-retrieval-benchmark-reread","2026-09-06T21:11:25+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"d4e4cbdc-ddde-458d-9649-e53bce5ddbce","MLPerf Training v6.0 把 MoE 钉在牌桌中央：671B 训练首次纳入工业基准，FP4 路径开始分裂","mlperf-training-v6-moe-671b-fp4-split","2026-06-16T18:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"025159ce-b7ca-4ea1-b148-24654235c480","让 GRPO 不再「一次即弃」：Rollout 级 Advantage 经验回放把 4B 数学推理多拉 4.35 pp","grpo-rollout-advantage-replay-4-35pp-math","2026-06-04T06:53:10+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"005557c5-8a3c-4d34-89bc-35d5351c4570","蒸馏只需要一条训练样本?清华实测:单条query覆盖71.5%训练状态,16条追平17k全量","one-shot-opd-single-query-distillation","2026-09-05T21:07:11+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"7623f190-7071-4811-a6f1-32462a99b8d3","经验会过期:阿里云论文让自主后训练的有害授权率从 62.5% 降到 25%","bcit-conditional-experience-transfer-post-training","2026-09-05T17:11:11+00:00",{"id":74,"title":75,"news_slug":76,"published_at":77},"00346b75-f071-42fd-ae16-db4c5569f01a","EarlyEval 提前叫停注定失败的 Agent:近半 token 省下,分辨率只动一两个点","earlyeval-early-stop-agent-eval","2026-09-03T21:04:52+00:00"]