Baidu released the downloadable weights for ERNIE-5.5-Flash-Base on November 5, 2026 — a 280-billion-parameter mixture-of-experts model that activates just 14 billion parameters per token. It is a raw pre-training checkpoint, not a finished chatbot, and that is the point: Baidu is handing developers a fixed starting point for their own fine-tuning instead of another closed API.
November 5, 2026
What Baidu released
The ERNIE team published the base checkpoint on November 5, ending at the pre-training stage — before the mid-training and post-training that turn a base model into an instruction-following assistant. Baidu positions it for domain adaptation, instruction tuning and reinforcement-learning research, and as a fixed baseline so different training methods can be compared against the same starting weights. A post-trained ERNIE 5.5 Flash model is promised to follow.
It is Baidu’s first open-weight release in roughly 16 months: the ERNIE 4.5 series was open-sourced in June 2025, while ERNIE 5.0 and 5.1 stayed closed. Downloads are available in FP8 and BF16 safetensors formats, and the model card documents serving with SGLang and vLLM.
Inside the model
ERNIE 5.5 Flash Base has 43 layers and was pre-trained on 33.5 trillion tokens using the Muon optimizer and FP8 mixed precision. Training context was gradually extended to 128K tokens, while fine-tuning and inference support up to 1 million tokens. Thirty-seven layers use a heavily compressed attention design that caches one entry per 128-token block, and six latent-attention layers keep token-level access to the full context. Its expert blocks route each token to 10 of 512 experts, plus one shared expert. The practical catch: even though only 14B parameters activate per token, the full 280B pool has to fit in memory, so this is a multi-GPU checkpoint, not a laptop model.
What Baidu’s benchmarks show
Baidu compared the checkpoint against the publicly released DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base using the same harness. In Baidu’s own results, ERNIE 5.5 Flash Base scored 93.3 on MMLU (five-shot) against DeepSeek’s 89.4, and 82.3 on HumanEval+ against 69.5. It did not sweep the table: Hy3-preview led on FullStackBench (56.7 to 54.6) and DeepSeek led on InfiniteBench (48.0 to 46.5). These are Baidu-reported figures and have not been independently reproduced.
Read the license before you download
Unlike ERNIE 4.5, which shipped under Apache 2.0, this release uses a custom license. Its territory excludes the European Union, the United Kingdom and South Korea. Non-commercial use is free of license charges; commercial use at or below $50 million in annual revenue from related products requires a one-time notice and prominent “Powered by ERNIE” attribution, and above that line Baidu’s prior written authorization is required. The license also bars using the model or its outputs to improve an unrelated AI system.
Why it matters
For builders outside the excluded regions, a competitive 280B base checkpoint shifts spending from pre-training — the most expensive step — to fine-tuning and serving, which smaller teams can actually afford. But the license is now part of the engineering decision: geography, revenue thresholds and the no-distillation clause will rule this model out for some teams before the benchmarks even matter. Everything AI, tested — and in this case, the fine print is half the story.
Frequently asked questions
Is ERNIE 5.5 Flash Base a finished chatbot model?
No. It is a base checkpoint that ends at pre-training, before instruction tuning. Baidu says a post-trained ERNIE 5.5 Flash will follow; the base is meant for teams doing their own fine-tuning and research.
How big is ERNIE 5.5 Flash Base?
It has 280 billion total parameters and activates 14 billion per token. It was pre-trained on 33.5 trillion tokens and supports up to 1 million tokens of context for fine-tuning and inference.
Can anyone use it commercially?
Not everywhere. The custom license excludes the EU, UK and South Korea, requires notice and attribution for commercial use up to $50 million in related annual revenue, and requires Baidu’s written authorization above that.
Sources: Unite.AI, KuCoin News, ChainCatcher, AlphaSignal, Runtime Wire