Baidu Releases 280B ERNIE 5.5 Flash Base: Restricted License Excludes EU and UK

Baidu released ERNIE-5.5-Flash-Base on November 5, its first open-weight model at scale since the ERNIE 4.5 family in June 2025. The checkpoint packs 280 billion total parameters into a sparse mixture-of-experts architecture that activates roughly 14 billion parameters per token during inference—delivering full-model expressiveness at a fraction of the per-token compute cost. The release targets researchers and enterprises seeking a large pre-trained foundation for domain adaptation rather than an out-of-the-box assistant.

What separates this release from ERNIE 4.5, however, is not its architecture but its license. The Apache 2.0 grant that covered ERNIE 4.5 is gone. In its place sits the ERNIE 5.5 Flash Base Model License Agreement v1.0—a custom document that bars commercial use by organizations operating in the European Union, the United Kingdom, and South Korea, institutes a revenue-tiered approval system, and explicitly prohibits using the weights to improve unrelated AI systems, a provision aimed at model distillation.

A Custom License Narrows the Model's Commercial Reach

The territorial restriction is the sharpest departure from ERNIE 4.5's terms. Under the new agreement, any company whose primary commercial operations fall in the EU, UK, or South Korea may not deploy the model commercially, regardless of where their inference infrastructure sits. For organizations below $50 million in annual revenue, commercial use requires a one-time notice to Baidu and attribution labeling reading "Powered by ERNIE." Above $50 million, Baidu's prior written authorization is required—a clause that gives the company visibility into and effective veto power over high-revenue deployments.

Baidu has not publicly explained the territorial carve-outs. The restrictions align with regulatory risk calculus common among Chinese technology companies: EU AI Act obligations, UK data-protection law, and South Korea's Personal Information Protection Act each impose compliance burdens that a one-sided model license cannot address. The restrictions may also reflect sensitivity around high-parameter model weights under evolving export-control frameworks.

Architecture: 43 Layers, 512 Experts, Two Distinct Attention Systems

ERNIE-5.5-Flash-Base runs on a 43-layer transformer with one dense feed-forward layer followed by 42 layers of Multi-Router Latent Mixture-of-Experts. Each MoE layer contains 512 routed experts, of which 10 are activated per token, plus one shared expert that always fires. Experts operate in a 2,048-dimensional latent space—half the model's 4,096-wide hidden dimension—cutting the compute cost of expert dispatch roughly in half compared with a full-width design.

The model uses two distinct attention mechanisms. Thirty-seven layers use Heavily Compressed Attention, which stores one KV cache entry per 128-token block plus a short recent window, dramatically cutting KV memory consumption for long sequences. The remaining six layers use Multi-head Latent Attention, which compresses keys and values into a low-rank representation and uses a learned indexer to retrieve high-scoring earlier positions. Together, these allow the model to handle its 128,000-token training context—extendable to one million tokens for fine-tuning—without proportionally scaling cache requirements.

Training used the Muon optimizer, a momentum-based alternative to AdamW adapted for distributed MoE training. Weights ship in FP8 precision (approximately 288 gigabytes) with a BF16 version also available.

Benchmark Results Are Baidu-Run Against Base Model Peers

Baidu's internal evaluation places ERNIE-5.5-Flash-Base ahead of comparable base models on 24 of 30 benchmarks. On MMLU (5-shot) it scores 93.3 against 89.4 for DeepSeek-V4-Flash-Base and 87.5 for Hunyuan3. HumanEval+ (1-shot Pass@1) reaches 82.3 versus 69.5 and 78.0 for the same comparators. MMLU-Pro comes in at 78.4 and GPQA at 60.2.

The evaluation is not a clean sweep. DeepSeek-V4-Flash-Base leads on InfiniteBench (48.0 versus 46.5) and Hunyuan3 leads on FullStackBench (56.7 versus 54.6). No independent evaluation has been published as of this writing. All numbers are company-reported, run by Baidu against base-model peers under conditions Baidu selected.

Deployment Requires Data-Center Hardware; Local Inference Is Not Yet Supported

Loading ERNIE-5.5-Flash-Base for inference requires at least 288 GB of GPU memory for the FP8 checkpoint. Supported frameworks are SGLang and vLLM together with Baidu's FlashMask attention kernels, which target NVIDIA Hopper and Blackwell GPUs. Baidu's recommended fine-tuning setup calls for 16 nodes of four GB300 GPUs each. As of publication, llama.cpp and mlx-lm do not support the architecture, making desktop or consumer-hardware inference unavailable.

Chinese Open-Weight Models Diverge on License Openness

ERNIE-5.5-Flash-Base arrives one day after Moonshot AI released weights for Kimi K3.1 Pro and three days after Meituan made LongCat-2.5 available via API. The closest architectural peer, DeepSeek-V4-Flash-Base, launched in April 2026 under an MIT license—the least restrictive option—with 284 billion total parameters and 13 billion active per token.

Where DeepSeek removed nearly every downstream restriction, Baidu is reasserting control over high-revenue commercial applications and establishing territorial exclusions that complicate adoption across three major markets. Organizations evaluating ERNIE 5.5 Flash Base as a fine-tuning foundation must weigh its technical profile against a bespoke license when MIT-licensed alternatives with similar parameter counts exist.

A post-trained ERNIE 5.5 Flash version is expected to follow. Whether that release carries the same territorial restrictions will determine how broadly the model is adopted for production AI stacks outside China.