百度发布 2800 亿参数开源基础模型 ERNIE 5.5 Flash Base
Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model
- 1百度发布开源基座模型 ERNIE-5.5-Flash-Base，不含微调与后训练，后训练版将随后推出
- 2模型采用 MoE 架构，总参数 2800 亿、每输入激活 140 亿，预训练数据 33.5 万亿 token
- 3训练上下文最高 128,000 token，设计上支持微调与推理阶段最长 100 万 token 上下文
- 4权重以 FP8 与 BF16 格式上架 Hugging Face，百度称基准表现与同规模开源模型相当
百度把一个 2800 亿参数的 MoE 基础模型直接开源，将后训练与场景适配的选择权和成本交给开发者。
- 影响：权重可自托管评估，每输入仅激活 140 亿参数有利于压低推理开销；基座版不带后训练，缺微调能力的团队需等官方后训练版本；百万 token 上下文设计对长文档场景有参考价值。
- 参考：可考虑先从 Hugging Face 下载 FP8 权重做小规模试点，用自有数据在推理、代码与长上下文任务上与 DeepSeek V4 Flash 等同规模模型对比，再决定投入。
- 依据：参数、训练数据与上下文规格为官方发布信息；与 DeepSeek V4 Flash、MiMo V2.5 的对比结论出自百度口径，暂无独立验证。
Baidu's AI development is still alive, but most of the work is on you. The Chinese company has released ERNIE-5.5-Flash-Base, an open foundation model that ships deliberately without fine-tuning or post-training. Developers are expected to decide for themselves where to take the model. A post-trained version is supposed to follow.
ERNIE 5.5 uses a mixture-of-experts architecture with 280 billion parameters, 14 billion of which are active per input. The model was pre-trained on 33.5 trillion tokens with a training context of up to 128,000 tokens, though it's designed to support a context of up to one million tokens during fine-tuning and inference.
In benchmarks for reasoning, code, and long-context tasks, the model competes with similarly sized open foundation models like DeepSeek V4 Flash or MiMo V2.5, according to Baidu. The weights are available on Hugging Face in FP8 and BF16.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
来源：The Decoder：AI News · the-decoder.com