Baidu opens 280B ERNIE base model; licence excludes EU, UK, Korea

Baidu has published ERNIE-5.5-Flash-Base on Hugging Face, a 280-billion-parameter mixture-of-experts language model released as a raw pretraining checkpoint, with a technical post and model card going live on November 5, 2026. The most important fact for anyone planning to use it is the licence: it is not Apache 2.0, and it does not extend to the European Union, the United Kingdom or South Korea.

This is also not a chat model. The weights are the end-of-pretraining checkpoint with no mid-training and no post-training applied, and Baidu’s release post says the post-trained ERNIE 5.5 Flash “will follow”.

What Baidu put on Hugging Face

The release consists of two repositories, baidu/ERNIE-5.5-Flash-Base in FP8 and baidu/ERNIE-5.5-Flash-Base-BF16, plus a technical post under Product Updates on Baidu’s ERNIE site. Both repositories are ungated, both are text-only, and both declare the same 280B total and 14B activated parameters.

The model is a sparse mixture-of-experts (MoE) design, meaning only a fraction of its parameters are used for each token. Baidu states 14B parameters are active per token, out of 280B in total, and that it was pretrained from scratch on 33.5 trillion tokens using the Muon optimizer and FP8 mixed precision.

The weights arrived well before the announcement. The Hugging Face repository was registered on October 21, 2026, and sat essentially empty until November 2, when configuration files landed at 06:29 followed by the weight shards; the last of the 128 FP8 shards finished uploading before 09:00 the next day. Benchmark and architecture figures went up on November 4, and the final README passes, tokenizer tweaks and logo came on November 5, within hours of the blog post. OrcaRouter, which tracked the commit history, summarises the order as weights first, documentation second, announcement third.

Early uptake is small. By the morning of November 5 the primary repository had been downloaded four times, against a like count in the high fifties.

ERNIE-5.5-Flash-Base specs and architecture

The model card is unusually specific about the architecture for a low-key release. The main figures, as reported:

Spec	Value
Total parameters	280B (Hugging Face widget: 286B)
Active parameters per token	14B
Pretraining data	33.5T tokens
Layers	43: one dense, 42 MoE
Experts	512 routed, top-10 routing, plus one shared expert
Hidden size / attention heads	4,096 / 64
Vocabulary	201,216 tokens
Context length	128K in training; released config enables 1,048,576 tokens for fine-tuning and inference
Formats	FP8 and BF16 safetensors, text-only

Attention is split between two mechanisms. Thirty-seven layers use a compressed attention Baidu calls HCA, which caches one entry per 128 tokens; the other six use MLA with DeepSeek-style sparse attention. Compressing the attention cache this way is what makes very long contexts cheaper to hold in memory, though the sources do not report measured memory or speed figures.

Why the parameter count differs between 280B and 286B

Two counts appear on the same page. Baidu’s 280B includes the input embedding and the output head, and its 14B activated figure counts the output head but not the embedding lookup; Hugging Face’s model-size widget reads 286B because it applies a different convention to the same files. Both figures exclude 0.36B of what Baidu calls ECHO parameters. For anyone sizing hardware, the practical point is that a spec sheet mixing the two conventions will not add up.

A base checkpoint, not a chat model

The model card describes the release as “the checkpoint at the end of pre-training, before any mid- or post-training”. Baidu’s stated purposes are domain adaptation, instruction tuning and reinforcement-learning research.

In practice, a base model continues text rather than following instructions or holding a conversation. The card serves it through the /v1/completions endpoint rather than chat completions, and notes that inputs need a BOS (beginning-of-sequence) token, so tokenisation has to keep add_special_tokens=True. Dropping it into an existing chat product will not produce a working assistant.

What did not ship, according to OrcaRouter: an instruct model, a usable chat template, an API listing, a price, or any independent benchmark. OrcaRouter also points out that this makes direct comparisons with finished, priced models such as Anthropic’s Claude Haiku 5.5 impossible for now, because one is a research artefact and the other is a hosted product.

The licence: regional exclusions and a revenue threshold

The licence is where this release differs most from a standard open-weight drop. According to OrcaRouter’s reading, it is not Apache 2.0, it does not extend to the EU, the UK or South Korea, and its commercial terms change once a user’s related revenue passes US$50 million a year. OrcaRouter names the licence, along with the base-only nature of the checkpoint, as the two properties of the release that explain the slow uptake.

That is a change from at least some of Baidu’s earlier ERNIE releases. The Hugging Face repository for ERNIE-4.5-0.3B-PT, created in June 2025, carries the Apache 2.0 licence, a permissive licence with no regional carve-outs. That small model has been downloaded 690,656 times in total, according to its repository metadata. OrcaRouter describes ERNIE-5.5-Flash-Base as Baidu’s second turn toward open weights.

The sources describe the headline terms but do not reproduce the full licence text, so details such as how “related revenue” is defined or what happens above the threshold are not covered here. Teams should read the licence file in the repository themselves before building on the weights.

Baidu’s benchmarks, and what is missing

The only scores that exist are Baidu’s own. The company ran 30 benchmarks under a unified harness against three other base models: DeepSeek-V4-Flash-Base (284B total, 13B active), Hy3-preview-Base (295B total, 21B active) and MiMo-V2.5-Base (310B total, 15B active). The sources do not report the individual results, so this article does not characterise how ERNIE-5.5-Flash-Base performed in that comparison.

The choice of comparison set is sensible on paper: all four are sparse MoE base models in a similar total-parameter range, with active counts between 13B and 21B. But base-model benchmarks run by the model’s maker under its own harness are a weak basis for picking a model. There is no Artificial Analysis page for the model yet and no listing in OrcaRouter’s 216-model catalogue as of its publication.

What it means for builders

The release is aimed at teams that train models, not teams that call them. Tooling for that shipped alongside: the model card carries SGLang and vLLM deployment examples, and NVIDIA’s Megatron-Bridge repository already has an ERNIE 5.5 example for fine-tuning. OrcaRouter calls this the practical footprint of a base release aimed at people who intend to train on it.

For builders, the practical considerations are:

• Jurisdiction first. Teams based in, or serving users in, the EU, the UK or South Korea are outside the licence’s scope as reported, which rules out the model regardless of its quality.
• Revenue planning. Companies whose related revenue could pass US$50 million a year face different commercial terms, which matters for anyone fine-tuning a product on top of it.
• Fine-tuning cost. With 280B total parameters, even with only 14B active per token, the full weights must be held in memory for training and serving. The FP8 repository reduces that footprint compared with BF16, but the sources give no hardware requirements.
• Long-context claims need testing. The model was trained at 128K context, while the released configuration enables 1,048,576 tokens. Whether quality holds at the longer setting is not shown in any source.
• Waiting may be simpler. Teams that want a usable ERNIE 5.5 Flash rather than a research base can wait for the post-trained version Baidu says will follow.

For teams that want a hosted Baidu model today, the company’s existing line continues through third-party providers. AIHubMix, for example, lists ERNIE 5.1, which it describes as the latest model in the Wenxin series, with a 119K-token context and text input and output.

What to watch next

The next development is the post-trained ERNIE 5.5 Flash that Baidu’s release post promises. Its licence will be as important as its scores: if it carries the same regional exclusions and revenue threshold as the base checkpoint, large parts of the developer market will remain outside it.

Independent evaluation is the other gap. Until third parties run their own tests, the 30-benchmark comparison against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base remains Baidu’s account alone. Download counts on the Hugging Face repositories, starting from four on launch morning, will show whether research teams outside the excluded regions take the base checkpoint up for fine-tuning.