Baidu Releases Weights for ERNIE 5.5 Flash Base, a 280-Billion-Parameter Base Model Whose License Excludes the European Union
On November 5, Baidu released ERNIE-5.5-Flash-Base, a 280-billion-parameter Mixture-of-Experts model (14 billion active) delivered straight out of pre-training, without chat tuning. Its benchmark scores surpass those of DeepSeek-V4-Flash-Base according to Baidu, but its proprietary license explicitly excludes the European Union, the United Kingdom, and South Korea.
Baidu released the weights for ERNIE-5.5-Flash-Base on November 5, 2026, a 280-billion-parameter Mixture-of-Experts (MoE) language model in which only 14 billion parameters are activated per token. A key aspect of this release: it is a base model, frozen at the end of pre-training, which Baidu presents as a starting point for research and fine-tuning. Furthermore, a specific clause in its license directly affects European readers: it excludes the European Union.
A Base Model, Not an Assistant
ERNIE-5.5-Flash-Base has received neither mid-training nor post-training: it does not follow instructions or engage in conversation; it performs text completion. Baidu intends it for domain adaptation, instruction tuning, and reinforcement learning research, promising that a post-trained version of ERNIE 5.5 Flash will “follow,” without offering a release date. This is an unusual strategy: most AI labs release the finished model first, with the base model provided as an afterthought.
The model was pre-trained from scratch on 33.5 trillion tokens, with a native context window of 128,000 tokens; Baidu states that fine-tuning and inference support up to one million tokens. The training corpus emphasizes repository-level code (commits, issues, pull requests, code reviews, and CI-validated bug fixes) as well as long-context comprehension. The weights are available in FP8 and BF16 on Hugging Face and ModelScope, accompanied by deployment guidelines for SGLang and vLLM.
An Architecture Designed for Cost Efficiency
Most of the technical post focuses on efficiency. Across the 43 layers, only one in seven retains token-by-token access to the entire context (via sparse MLA latent attention); the other 37 layers rely on heavily compressed attention that preserves just one cache entry per 128 tokens. The goal is to slash the memory footprint required for long contexts, a major cost driver for agentic workflows. Each MoE layer includes 512 experts (10 active per token) plus one shared expert, selected by multiple independent routers rather than a single routing mechanism. Baidu also replaces the multi-token prediction module with the network’s final layer, which doubles as a draft model for speculative decoding.
Outperforming DeepSeek, According to Baidu
Baidu compares its model to three Chinese base models of similar size, evaluated using what it claims is the same protocol: DeepSeek-V4-Flash-Base (284 billion parameters, 13 active), Hy3-preview-Base (295 billion parameters), and MiMo-V2.5-Base (310 billion). ERNIE comes out ahead on most benchmarks: 78.4 on MMLU-Pro compared to 69.2 for DeepSeek, 82.3 versus 69.5 on HumanEval+, and 40.1 versus 31.9 on LiveCodeBench. Conversely, it trails on ChineseSimpleQA (73.0 vs. 75.3 for DeepSeek), FullStackBench (54.6 vs. 56.7 for Hy3-preview-Base), and InfiniteBench. These metrics come directly from the developer and reflect a base model: they offer no guarantees about the performance of the upcoming conversational version.
A Highly Restrictive Custom License
Baidu’s custom license is far from open source. It outlines an authorized territory covering the entire world except the European Union, the United Kingdom, and South Korea, specifying that the model, its derivatives, and even its outputs must not be used, reproduced, or distributed outside this designated area. Elsewhere, non-commercial use is unrestricted; commercial use is free as long as the products involved do not exceed $50 million in annual revenue, provided users submit a one-time notification to Baidu and display a “Powered by ERNIE” badge. Beyond that threshold, prior written authorization is required. The license also explicitly bars using the model or its outputs to improve other AI models besides its own derivatives, as well as any military use.
Our take
Technically speaking, ERNIE-5.5-Flash-Base is a substantial release: a modern base model with a reasonable active parameter count, detailed architecture documentation, and a genuine push to reduce long-context operational costs. Publishing an “untouched” checkpoint is also a welcome move for researchers, who often lack clean foundation models to benchmark different post-training techniques. However, for a European audience, the main takeaway lies elsewhere: the license explicitly excludes the European Union, and Baidu offers no explanation for this decision in its materials. Companies and labs based in France or across the EU are therefore barred from using it, and the restriction against using outputs to train other models further diminishes its appeal elsewhere. The contrast is striking with the much more permissive terms of Kimi K3.1 Pro, which we covered yesterday: the open-weights race is now being fought on terms of service, not just benchmark scores.
Sources
- Baidu — ERNIE 5.5 Flash Base, technical post (November 5, 2026) Primary source
- Baidu — ERNIE-5.5-Flash-Base model card (Hugging Face) Primary source
- Baidu — ERNIE 5.5 Flash Base Model License Agreement v1.0 (Hugging Face) Primary source
- Baidu — BF16 version ERNIE-5.5-Flash-Base-BF16 (Hugging Face) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
