Baidu Inc. released its latest foundation artificial intelligence model under an open-source license on Friday, marking a decisive strategic pivot for a company whose chief executive officer previously dismissed open-source software as a commercial dead end.
The Chinese technology giant made weights for its ERNIE 5.5 Flash Base model publicly available on Hugging Face in FP8 and BF16 precision formats. The move comes after Baidu Chief Executive Officer Robin Li repeatedly described open-source AI as a "wealth trap," maintaining that closed-source proprietary systems offered the only viable commercial path.
Industry observers view the release as a direct response to shifting market dynamics and intense domestic competition, largely catalyzed by the popularity of open-source models from rivals such as DeepSeek. Baidu began revising its stance last June by opening access to its ERNIE 4.5 model before releasing the upgraded ERNIE 5.5 Flash Base architecture.
The newly released model utilizes a Mixture of Experts (MoE) design containing 280 billion total parameters. To maintain computational efficiency, the system activates only 14 billion parameters per token during processing.
Baidu pre-trained the model on 33.5 trillion tokens of data. The architecture natively supports a 128,000-token context window during pre-training, which expands up to 1 million tokens during inference after domain-specific fine-tuning, enabling the processing of extensive documents and entire software repositories.
Engineering optimizations include a hybrid sparse attention mechanism to compress historical context and minimize GPU memory usage during long-text tasks. Baidu also incorporated a multi-router expert mechanism, the Muon optimization algorithm, FP8 mixed-precision training, and a proprietary attention operator designated FlashMask-4.
To facilitate rapid enterprise adoption, Baidu enabled out-of-the-box compatibility with mainstream open-source inference frameworks, including vLLM and SGLang.
In benchmark evaluations released by the company, ERNIE 5.5 Flash Base scored 93.3 on MMLU and 78.4 on MMLU-Pro, leading competing base models of similar active parameter sizes. On programming benchmarks, the model registered 82.3 on HumanEval+ and 81.5 on MBPP+, outperforming comparable models such as DeepSeek-V4-Flash-Base, though analysts noted it slightly lagged DeepSeek on the SimpleQA factuality test.
Baidu clarified that the current download represents a raw pre-trained base model intended for developer fine-tuning, domain adaptation, and reinforcement learning. A fully post-trained instruction version, ERNIE 5.5 Flash, will be released in a subsequent update.
