In Strategic Reversal, Baidu Open-Sources 280-Billion Parameter ERNIE 5.5 AI Mannequin
Baidu Inc. launched its newest basis synthetic intelligence mannequin below an open-source license on Friday, marking a decisive strategic pivot for an organization whose chief govt officer beforehand dismissed open-source software program as a industrial lifeless finish.
The Chinese language expertise large made weights for its ERNIE 5.5 Flash Base mannequin publicly accessible on Hugging Face in FP8 and BF16 precision codecs. The transfer comes after Baidu Chief Govt Officer Robin Li repeatedly described open-source AI as a “wealth lure,” sustaining that closed-source proprietary programs provided the one viable industrial path.
Business observers view the discharge as a direct response to shifting market dynamics and intense home competitors, largely catalyzed by the recognition of open-source fashions from rivals akin to DeepSeek. Baidu started revising its stance final June by opening entry to its ERNIE 4.5 mannequin earlier than releasing the upgraded ERNIE 5.5 Flash Base structure.
The newly launched mannequin makes use of a Combination of Specialists (MoE) design containing 280 billion complete parameters. To take care of computational effectivity, the system prompts solely 14 billion parameters per token throughout processing.
Baidu pre-trained the mannequin on 33.5 trillion tokens of knowledge. The structure natively helps a 128,000-token context window throughout pre-training, which expands as much as 1 million tokens throughout inference after domain-specific fine-tuning, enabling the processing of intensive paperwork and whole software program repositories.
Engineering optimizations embrace a hybrid sparse consideration mechanism to compress historic context and decrease GPU reminiscence utilization throughout long-text duties. Baidu additionally integrated a multi-router knowledgeable mechanism, the Muon optimization algorithm, FP8 mixed-precision coaching, and a proprietary consideration operator designated FlashMask-4.
To facilitate speedy enterprise adoption, Baidu enabled out-of-the-box compatibility with mainstream open-source inference frameworks, together with vLLM and SGLang.
In benchmark evaluations launched by the corporate, ERNIE 5.5 Flash Base scored 93.3 on MMLU and 78.4 on MMLU-Professional, main competing base fashions of comparable lively parameter sizes. On programming benchmarks, the mannequin registered 82.3 on HumanEval+ and 81.5 on MBPP+, outperforming comparable fashions akin to DeepSeek-V4-Flash-Base, although analysts famous it barely lagged DeepSeek on the SimpleQA factuality take a look at.
Baidu clarified that the present obtain represents a uncooked pre-trained base mannequin meant for developer fine-tuning, area adaptation, and reinforcement studying. A totally post-trained instruction model, ERNIE 5.5 Flash, might be launched in a subsequent replace.
