Baidu releases ERNIE 5.5 Flash Preview
Baidu released ERNIE 5.5 Flash Preview through the Qianfan API, targeting programming, 3D modeling, knowledge work and code-based content creation. Baidu says the sparse MoE model has 280 billion total parameters, activates 14 billion per token, supports a 1 million-token context and exceeds 200 tokens per second on the Qianfan API. PinWest’s testing found strong frontend and code-animation results, but recurring errors in unconstrained complex 3D structures and physical-collision scenes; Baidu plans to release the weights under Apache 2.0 by November 25.
The model combines a long context, vendor-reported API speed and low per-token pricing with planned open weights, while independent testing suggests it may work well for coding workflows that use automated checks and multiple correction rounds.
这周的大模型圈格外拥挤,Claude Fable 5.2、Gemini 4 Carbon、Qwen 4 Max Preview 前后脚更新,各家都在集中展示大参数旗舰模型的上限。
百度在自己的世界大会前夕开启了一轮密集更新,11 月 5 日开源基座模型文心5.5 Flash Base,11 月 9 日披露参与新模型研发的自动研究系统 ERNIE Evolve,11 月 11 日正式发布主打速度的文心5.5 Flash Preview(ERNIE 5.5 Flash Preview),把重心放在了编程、3D 建模、知识工作与基于代码的内容创作上。
从参数设计上看,这款模型做了明确取舍。它采用稀疏 MoE 架构,总参数量达 2800 亿,每个 Token 激活 140 亿参数,支持最长 100 万 Token 的上下文。无论是后训练阶段的反复采样,还是 Agent 任务里的多步交互与长文本处理,算力和显存开销都会持续累加。把激活规模控制在 140 亿,换来的是更轻的推理负担。官方给出的千帆 API 单用户生成速度超过 200 tokens/s,定价为每百万 Token 输入忙时 2 元、闲时 1 元,输出忙时 8 元、闲时 4 元,限时五折。