Baidu has released the base weights for ERNIE-5.5-Flash-Base, a sparse mixture-of-experts model with 280 billion total parameters and about 14 billion activated per token, the company said in a 5 November ERNIE blog post. The checkpoint is also listed on Hugging Face as baidu/ERNIE-5.5-Flash-Base.
Baidu describes the drop as an end-of-pre-training checkpoint, before mid-training or post-training. A post-trained ERNIE 5.5 Flash “will follow,” the blog says. That makes the release a research and adaptation starting point, not a finished chat model.
In brief
- Open weights: 280B total / ~14B active MoE; 43 layers; trained from scratch on 33.5T tokens.
- Context: 128K in training; up to 1M for fine-tuning and inference, Baidu says.
- Licence territory excludes the EU, UK and South Korea; say open weights, not open source.
Architecture and Baidu’s own scores
According to the blog, the model uses hybrid multi-head latent attention with heavily compressed attention (one full-context layer in seven), 512 routed experts with top-10 routing plus one shared expert, Muon optimisation and FP8 mixed precision in pre-training. Training context reached 128K tokens; Baidu says fine-tuning and inference can use up to 1M tokens.
Baidu published its own evaluation table comparing the base checkpoint with DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base. On that table, ERNIE-5.5-Flash-Base scores 78.4 on MMLU-Pro versus 69.2 / 66.9 / 65.7 for those three, and 40.1 on LiveCodeBench versus 31.9 / 33.5 / 32.4. Those figures are Baidu-reported; TSN has not re-run the evals.
The move sits alongside other Chinese and open-weights drops TSN has tracked, from Baidu’s Unlimited-OCR 2 to Ricursive’s Ripple-65B preview and Deep Principle’s Hydrogen-1.0 science model. The common thread is shipping a base checkpoint early so outsiders can fine-tune before the vendor’s own post-trained flash variant arrives.
Licence: weights, not open source
The Hugging Face LICENSE file for the model is the ERNIE 5.5 Flash Base Model License v1.0. “Territory” means worldwide areas other than the European Union, the United Kingdom and South Korea. Use, distribution and display outside that territory are outside the grant. Non-commercial research use is free within the territory; commercial use needs a one-time notice below a $50 million related-revenue threshold and Baidu’s separate prior written authorisation above it.
That is why TSN calls the release open weights rather than open source. OSI-style open source would not carve the UK and EU out of the map or gate high-revenue commercial use behind a vendor authorisation email. UK builders reading the card should treat the weights as unavailable under this licence, not as a free local download target.
Hybrid attention and a research-first checkpoint
Baidu’s blog leans on the idea that a clean pre-training end-state is more useful to outsiders than another instruct-tuned flash model nobody can re-base. The hybrid attention recipe — full context on only one layer in seven, with compressed attention elsewhere — is presented as the way to keep long-context training affordable at 280B total scale. Readers should still treat every leaderboard cell as Baidu’s own methodology until third parties reproduce it.
Why a UK-excluding base still matters to labs elsewhere
For labs inside the permitted territory, an end-of-pre-training 280B/14B MoE with a promised post-trained sibling is a practical fork point: adapt the base for domain work now, or wait for Baidu’s flash instruct model. For UK and EU teams, the story is different — the interesting signal is geopolitical as much as technical. Baidu is willing to publish a large base checkpoint while writing the UK and EU out of the licence territory, which narrows who can legally productise the weights even if the files are visible on Hugging Face. Visibility on a model hub is not the same as a licence to fine-tune for a London or Berlin product.
That carve-out also reshapes competitive comparisons. Western open-weights drops that still permit UK and EU commercial use will look more attractive to regulated buyers even when Baidu’s reported MMLU-Pro number looks higher on paper. Procurement and counsel will ask about territory before they ask about LiveCodeBench.
Open questions are when the post-trained ERNIE 5.5 Flash arrives, whether Baidu publishes training code or only weights, and how independent evals compare once outsiders can run them inside the territory. Until then, the confirmed facts are the architecture numbers, the Baidu-reported benchmark table, and a licence that is explicit about who is outside the map.
Sources
- ERNIE-5.5-Flash-Base release (Baidu ERNIE blog, 5 November 2026)
- baidu/ERNIE-5.5-Flash-Base model card (Hugging Face)
- ERNIE 5.5 Flash Base Model License v1.0 (Hugging Face LICENSE)
Related stories
- Baidu Releases Unlimited-OCR 2 for Documents Hundreds of Pages Long
- Ricursive Releases Ripple-65B Preview Open Weights
- Deep Principle Opens Hydrogen-1.0, a 753B-Parameter Science Model
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
