Blog · AI & automation

Baidu Opens ERNIE 5.5 Flash Base at 280B MoE

Share: https://www.ellixorlabs.com/blog/baidu-ernie-5-5-flash-base-280b-open-weights (copy URL for iMessage or social)

Published (display in IST)

Updated (IST)

Close-up of a server unit in a blue-lit data centre
Photo: panumas nikhomkhai / Pexels (https://www.pexels.com/photo/computer-server-in-data-center-room-17489163/)

Baidu released ERNIE 5.5 Flash Base, a 280B MoE pre-training checkpoint with 14B active parameters, company-reported scores versus peer base models, and a licence that excludes the EU, UK and South Korea.

Photo: panumas nikhomkhai / Pexels

Baidu has open-weighted ERNIE 5.5 Flash Base, a sparse Mixture-of-Experts language model with 280 billion total parameters and 14 billion activated per token. The 5 November 2026 release is the raw pre-training checkpoint, before mid- or post-training, aimed at teams that want a shared starting point for domain adaptation, instruction tuning and reinforcement-learning research.

What happened

According to Baidu's official ERNIE blog and the Hugging Face model card, ERNIE-5.5-Flash-Base was pre-trained from scratch on 33.5 trillion tokens with a training context of up to 128K tokens, while fine-tuning and inference support a 1M-token context. Architecture highlights include hybrid Heavily Compressed Attention and Multi-head Latent Attention across 43 layers, Multi-Router Latent MoE with 512 routed experts (top-10) plus one shared expert, and ECHO, which reuses the backbone's final layer instead of a separate multi-token prediction module. Weights are available in FP8 and BF16 on Hugging Face and ModelScope, with documented serving paths through SGLang and vLLM. Baidu's own unified evaluation, which it says used the same harness, prompts and few-shot settings for all models, reports ERNIE ahead of DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base on many tasks: MMLU-Pro 78.4 versus 69.2, 66.9 and 65.7; HumanEval+ 82.3 versus 69.5, 78.0 and 74.4; LiveCodeBench 40.1 versus 31.9, 33.5 and 32.4; LongBench-v2 48.7 versus 44.5, 39.8 and 44.6. Treat every figure as company-reported until independent benches such as LMArena or Artificial Analysis publish their own runs. MarkTechPost summarised the same numbers and noted losses on a minority of tasks, including HELMET-Rerank behind DeepSeek-V4-Flash-Base.

Why it matters for founders

For founders, the story is not another chat model. It is an open base checkpoint you can fine-tune, with efficiency choices aimed at long contexts and further training cost. That matters if you are building specialist agents or domain models and want a competitive MoE starting point without owning a full pre-training run. Read the licence before you plan a product: Territory excludes the European Union, the United Kingdom and South Korea, so use, distribution and hosting outside those limits are not granted. Commercial use at or below US$50 million related annual revenue is free with "Powered by ERNIE" attribution and a one-time notice to Baidu; above that threshold you need prior written authorisation. Indian and other permitted-market teams still need to weigh hosting location, customer geography and derivative-model rules carefully. Teams shipping agents should also keep watching how open bases compare with closed frontier stacks covered in our Thinking Machines Inkling 2 Small note and Gemini Omni 1.2 Flash pricing update, and how security-focused specialisation is evolving in OpenAI's GPT-6 Cyber launch. If you need help turning open weights into a production agent stack, our AI agents service maps evaluation, tooling and delivery choices.

What to watch next

Watch for Baidu's promised post-trained ERNIE 5.5 Flash, independent leaderboard runs, and how the territorial licence shapes adoption in India and other permitted markets versus EU and UK peers. Primary sources: Baidu ERNIE blog at https://ernie.baidu.com/release/ernie-5-5-flash-base/, Hugging Face card at https://huggingface.co/baidu/ERNIE-5.5-Flash-Base, and MarkTechPost summary at https://www.marktechpost.com/2026/11/05/baidu-ernie-5-5-flash-base-280b-moe/.

Frequently asked questions

What is ERNIE 5.5 Flash Base? It is Baidu's 280B total, 14B active MoE pre-training checkpoint released on 5 November 2026, not a finished chat model. Are the benchmark scores independent? No. Baidu reports them from its own unified pipeline against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base; label them company-reported until third parties replicate. Can UK or EU teams use the weights? The licence Territory excludes the European Union, the United Kingdom and South Korea, so those uses are outside the rights granted.

If you are evaluating open base models for product agents or domain fine-tunes and want a clear build plan, talk to us at https://www.ellixorlabs.com/contact or explore https://www.ellixorlabs.com/services/ai-agents.

FAQ

What is Baidu's ERNIE 5.5 Flash Base model?

ERNIE 5.5 Flash Base is a 280 billion parameter sparse MoE language model from Baidu with 14 billion parameters activated per token. Released on 5 November 2026, it is the raw pre-training checkpoint before mid- or post-training, intended as a starting point for further training.

Are ERNIE 5.5 Flash Base benchmark scores independent?

No. The MMLU-Pro, HumanEval+ and other figures cited in coverage come from Baidu's own unified evaluation against peer base models. Treat them as company-reported until independent benches publish matching runs.

Can UK or EU companies use ERNIE 5.5 Flash Base weights?

The ERNIE 5.5 Flash Base Model License Agreement defines Territory as worldwide areas other than the European Union, the United Kingdom and South Korea. Use and distribution outside that Territory are not authorised under the licence.

Talk to Ellixor Labs about this topic.

Source: Baidu ERNIE