Baidu open-sources 280B parameter ERNIE 5.5 Flash base model

Baidu just dropped its first open model in 16 months, giving developers a free 280-billion-parameter base model to download and fine-tune.

Models · Source: Unite.AI

What happened

Baidu just released ERNIE 5.5 Flash Base. It is a text-only mixture-of-experts model with 280 billion total parameters. Only 14 billion parameters activate per token. This makes it highly efficient for its massive size. It is Baidu's first open-source model release in 16 months, breaking a silence that lasted through their 5.0 and 5.1 versions. Developers can download the weights directly in FP8 and BF16 formats today.

The model architecture is built for scale. It features 43 layers and 512 routed experts. Baidu trained it from scratch on 33.5 trillion tokens. It supports a massive one-million-token context window for fine-tuning and inference. Baidu achieved this by compressing historical tokens into single cache records across most of its layers, keeping memory costs down as the context grows. They also released a FlashMask repository to optimize attention kernels for Nvidia Hopper and Blackwell GPUs.

Baidu claims the model dominates the competition. In their internal evaluations, ERNIE 5.5 Flash Base beat DeepSeek V4 Flash Base, Tencent Hy3 preview Base, and Xiaomi MiMo V2.5 Base in 24 out of 30 benchmarks. It scored 93.3 on MMLU and 82.3 on HumanEval+. But remember this is a raw base model. It has not undergone instruction tuning or reinforcement learning. It cannot act as a chatbot out of the box. A post-trained version will follow later.

Key facts

Why it matters

Builders now have a frontier-level starting point they can run on their own hardware. You do not have to pay API fees or worry about a provider shutting off your access. The mixture-of-experts design means you get the reasoning power of a 280-billion-parameter model but only pay the compute cost of a 14-billion-parameter model during inference. Because it is a raw base model, you have complete control over the domain adaptation. You can shape it specifically for agentic software engineering, repository-level reasoning, or long-document analysis using the massive one-million-token context window.

The catch is the restrictive new license. Baidu abandoned the open Apache 2.0 protocol it used for ERNIE 4.5. The new custom license strictly bans use, modification, or distribution in the European Union, the United Kingdom, and South Korea. It also explicitly forbids using the model or its outputs to train other AI systems. If your product makes over 50 million dollars a year, you lose the free commercial license. You must get Baidu's written permission to continue operating. This creates massive legal liabilities for fast-growing startups.

For builders

Build custom agents without API dependency

You get a massive base model you can run locally or on your own cloud infrastructure. You pay for the compute, but you avoid recurring API bills and vendor lock-in. It is a blank slate ready for your specific instruction tuning.

Beware the strict geographic and commercial limits

Founders targeting Europe, the UK, or South Korea cannot legally use this model. Startups crossing 50 million dollars in annual revenue lose their free commercial license and must negotiate directly with Baidu to survive.

No synthetic data generation for other models

The license strictly bans using ERNIE 5.5 Flash outputs to train or improve unrelated AI systems. Builders relying on distillation workflows must look elsewhere or risk severe legal action from Baidu.

My take

Baidu is giving us a massive, highly capable base model for free, but they are wrapping it in red tape. The geographic bans and anti-distillation clauses show they are terrified of competitors using their work. It is a fantastic sandbox for US developers, but do not build your core business on it if you plan to scale globally.

Original reporting: Unite.AI. This is my rewrite and opinion.

More AI news for builders