ERNIE 5.5 Flash Preview 可独立完成 RSI 讲解视频生成
Baidu 这个 280B 总参数、14B 激活的模型，8 分钟独立做完一条讲解视频，脚本、图表代码、渲染全包，比前沿模型快 2-5 倍。现在 OpenRouter 免费可试，11 月 25 日前还会开源权重，做 agent 工作流的人可以上手测测长视频连贯性到底行不行。
Baidu 发布 ERNIE 5.5 Flash Preview，参数规模 280B，激活参数 14B，上下文窗口 1M token。测试者只给出一个任务——解释递归自我改进（RSI）的短视频，模型独立运行约 8 分钟完成全部工作，包括编写脚本、为循环图和图表写代码、渲染最终成片。
工作流中 pi 作为 agent harness，ElevenLabs 提供语音，其余全部由 ERNIE 自主完成。据测试者对比，同样的任务用前沿模型需要 2 到 5 倍时间。Baidu 还针对 Unreal、Unity、Godot 等游戏引擎做了专门调优，知识工作场景下可规划任务并产出表格、文档、幻灯片和研究报告。
该模型目前可在 OpenRouter 免费使用，Baidu 承诺开放权重在 11 月 25 日前发布。对开发者而言，小激活参数加长 agent 运行时的组合，正在成为 Baidu 对抗 OpenAI、Google flash 级模型差异化的方向，开放权重落地后可自行部署验证。
This short RSI explainer was fully generated by a model. The model is ERNIE 5.5 Flash Preview, Baidu's new model. I gave it one task. "Make a short video that explains recursive self-improvement (RSI)." I used pi as the agent harness and ElevenLabs for the voice. Everything else came from ERNIE, working on its own for about 8 minutes. For the same video, it would take 2x-5x longer with frontier models. I am learning that these flash models can execute these tasks faster and with higher quality than was possible just a couple of months ago. It wrote the code for the loop diagram and the charts. It wrote the script and rendered the final film. This is the long-running agent work Baidu built it for. It's a 280B model with 14B active parameters and a 1M-token context window. I am not testing it with longer videos where coherence matters. Baidu also tuned it for game development in engines like Unreal, Unity, and Godot. For knowledge work, it can plan a task and produce spreadsheets, documents, slide decks, and research reports. It's free on OpenRouter right now. Baidu says the open weights will land before Nov 25. openrouter.ai/baidu/ernie-5.…
本条情报归属 AITOP「模型」核心栏目，重点聚焦中国 AI 产业与国产大模型生态。在跨区域产业观察坐标中，该动态直接关联「中美大模型技术代差与商业化对照」「AI 开发者工具链、Coding Agent 与工程化实践」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：ERNIE、Baidu、OpenRouter、智能体，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明：本文原始线索采编自「elvis」（@omarsar0），经 AITOP 多维评分管线结构化提炼。
