Editorial illustration for Baidu Releases Open-Source ERNIE 5.5 Flash Base With 280 Billion Parameters
Baidu Releases ERNIE 5.5 Flash: 280B Parameter Model
Baidu Releases Open-Source ERNIE 5.5 Flash Base With 280 Billion Parameters
Baidu put a new large language model on Hugging Face this week, and it comes with an unusual catch: nobody at the company bothered to fine-tune it. ERNIE-5.5-Flash-Base is a raw foundation model, 280 billion parameters built on a mixture-of-experts design, with only 14 billion active per input. Baidu trained it on 33.5 trillion tokens and gave it a 128,000-token training context, though the architecture is built to stretch to a million tokens once developers start fine-tuning or running inference.
The release pattern matters here. Rather than shipping something polished and ready to chat, Baidu is handing over a base model and leaving the post-training work to whoever downloads it. A fine-tuned version is expected eventually, but for now the burden sits with developers. Weights are up in both FP8 and BF16 formats, so there's flexibility on the deployment side at least.
Baidu is positioning ERNIE 5.5 against other open foundation models in its weight class, pointing to benchmark results in reasoning, code, and long-context handling. The company named specific rivals in that comparison, which gives some sense of where it thinks this model stands in the current open-source field.
Baidu's AI development is still alive, but most of the work is on you. The Chinese company has released ERNIE-5.5-Flash-Base, an open foundation model that ships deliberately without fine-tuning or post-training. Developers are expected to decide for themselves where to take the model.
Why this matters
Releasing a base model without post-training is a specific bet: Baidu is handing over a 280-billion-parameter MoE backbone, trained on 33.5 trillion tokens, and trusting the open-source community to do the alignment work it usually keeps in-house. For developers, that's an invitation to build custom fine-tunes rather than inherit someone else's safety tuning and refusal patterns, which cuts both ways. You get more control over behavior; you also inherit more responsibility for making the thing usable and safe before shipping it.
With only 14 billion of those 280 billion parameters active per input, the efficiency math looks attractive for teams without hyperscaler budgets, especially with a 128,000-token context window on offer. Researchers get a rare look at a frontier-scale Chinese model in its raw state, before any post-training layer smooths it over. The real test comes when Baidu's promised post-trained version lands: that will show whether this base was strong enough to be worth the extra work, or whether Baidu simply moved the hard part downstream and called it open source.
Related Reading
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Axiom Math raises USD 200M Series A led by Menlo to keep AI slop out of math
Related Topic Baidu launches ERNIE 5.0 preview 1022, beating GPT-5 on charts and document tasks
Business & Startups
Related Topic Tencent, Baidu launch AI companion apps; women dominate market, says Sun Zhaozhi
LLMs & Generative AI
From Archives Chris Alexiuk Makes AI Easy for Developers with Nemotron Open Models
From Archives Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active
From Archives Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing
From Archives Bernie Sanders proposes USD 7 trillion plan for U.S. public control of AI
Business & Startups
Common Questions Answered
What makes ERNIE-5.5-Flash-Base different from other large language models in terms of its release?
ERNIE-5.5-Flash-Base is released as a raw foundation model without any fine-tuning or post-training applied by Baidu. This means developers receive the base model directly and are responsible for their own alignment work and customization, rather than inheriting Baidu's pre-configured safety tuning and refusal patterns.
How many parameters does ERNIE-5.5-Flash-Base have and what is its architecture design?
ERNIE-5.5-Flash-Base contains 280 billion parameters built on a mixture-of-experts (MoE) design, with only 14 billion parameters active per input. This efficient architecture allows the model to maintain high performance while reducing computational requirements during inference.
What training data and context window was ERNIE-5.5-Flash-Base trained on?
Baidu trained ERNIE-5.5-Flash-Base on 33.5 trillion tokens with a 128,000-token training context window. The model's architecture is designed to support expansion up to a million-token context once developers begin fine-tuning or running inference on it.
What are the advantages and responsibilities for developers using ERNIE-5.5-Flash-Base?
Developers gain more control over the model's behavior by creating custom fine-tunes tailored to their specific needs, rather than being constrained by Baidu's pre-configured safety measures. However, this freedom also means developers inherit full responsibility for ensuring appropriate alignment and safety in their implementations.
Further Reading
- Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model - The Decoder
- Baidu Releases ERNIE 5.5 Flash Base Pretraining Weights - HyperAI
- ERNIE 5.5 Flash Base - Baidu ERNIE
- baidu/ERNIE-5.5-Flash-Base - Hugging Face
