Reportedly: Baidu open-sources ERNIE-5.5-Flash-Base, a 280B-parameter MoE model with 14B active parameters
Reported by Zhidongxi · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Baidu released and open-sourced ERNIE-5.5-Flash-Base, a sparse MoE model with 280B total and 14B active parameters, trained on 33.5T tokens. It supports 128K training context and 1M inference context. Weights are available on Hugging Face in FP8 and BF16, Zhidongxi reported.
China context
- Original name
- 文心5.5 Flash Base
- Outside China
- Open weights · zhidx.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the model weights from Hugging Face and integrate with vLLM or SGLang, enabling local deployment and fine-tuning without API access to Baidu's services.
- For investors
- The release intensifies competition in the open-weights segment, potentially pressuring valuations of startups whose models are benchmarked against ERNIE-5.5-Flash-Base.
Translated from Chinese. Quotes and facts link to the original sources.
Baidu has officially released and open-sourced ERNIE-5.5-Flash-Base, a sparse mixture-of-experts (MoE) base model with 280 billion total parameters and 14 billion active parameters per token. The model was pre-trained from scratch on 33.5 trillion tokens, with a maximum training context length of 128K and support for 1M-token context during fine-tuning and inference. According to Baidu's published evaluation results, the model achieves state-of-the-art performance among models of comparable size, particularly in reasoning, coding, academic knowledge, and long-context tasks, outperforming models such as DeepSeek-V4-Flash-Base (284B-A13B), Hy3-preview-Base (295B-A21B), and MiMo-V2.5-Base (310B-A15B). The model weights are openly available on Hugging Face in FP8 and BF16 formats, and Baidu has partnered with vLLM and SGLang to provide open-source ecosystem support.
The model uses a sparse MoE architecture with 43 Transformer layers, hidden size 4096, and 64 attention heads. Baidu emphasizes efficiency through architectural choices, expert routing, and next-token prediction optimizations to maximize activated computation and KV cache utilization. The 14B active parameters per token aim to balance model capacity and inference cost.
Baidu's open-weight release gives developers outside China direct access to a 280B-parameter MoE base model with 14B active parameters, reducing the cost of building on a state-of-the-art Chinese foundation model. Competitors offering similarly sized open models must now match or exceed its reported benchmark performance to retain developer mindshare.
Open-sourcing a high-performance base model can attract developers and enterprises to Baidu's ecosystem, potentially driving adoption of its commercial cloud and AI services. The release also positions Baidu competitively in the open-weights market against rivals like DeepSeek and Alibaba.
The next verifiable signal is whether independent benchmarks confirm Baidu's reported SOTA claims and whether the model gains adoption in downstream fine-tuning projects. Watch for community forks, fine-tuned variants, and integration into major inference frameworks beyond vLLM and SGLang.