Baidu releases ERNIE 5.5 Flash Base, a 280-billion-parameter open-weight model with a restrictive licence
Baidu has published ERNIE 5.5 Flash Base, a 280B-parameter mixture-of-experts checkpoint with 14B active per token, trained on 33.5 trillion tokens. Weights are downloadable, but the licence excludes the EU, UK and South Korea and limits commercial use above $50 million revenue.
Baidu's ERNIE team announced ERNIE 5.5 Flash Base on 5 November, releasing downloadable weights for a text-only mixture-of-experts model. It has 280 billion parameters in total, with 14 billion active per token, across 43 layers and 512 routed experts, ten of which fire per token, plus one shared expert.
The base checkpoint was released after pre-training but before mid-training or post-training, so it is a starting point for fine-tuning and research rather than a finished assistant. It was trained on 33.5 trillion tokens using the Muon optimiser and FP8 mixed precision, with training context extended to 128K tokens and support for up to one million tokens in fine-tuning and inference.
On Baidu's own evaluation it scores 93.3 on five-shot MMLU against 89.4 for DeepSeek V4 Flash, and 82.3 on HumanEval+ against 69.5. It trails on some tests: DeepSeek V4 Flash leads on InfiniteBench, and Tencent's Hy3 preview leads on FullStackBench. These are vendor figures and have not been independently verified.
The licence is the catch. Its territory excludes the European Union, the United Kingdom and South Korea. Non-commercial use carries no charge, commercial users with related annual revenue of $50 million or less must send a one-time notice and credit "Powered by ERNIE", and larger users need Baidu's written authorisation. Using the model or its outputs to improve unrelated AI systems is restricted.
Why it matters
Open weights no longer imply open terms. Territorial exclusions and revenue thresholds make legal review part of any model selection.
The release adds to a crowded field of large Chinese open-weight models competing on efficiency per active parameter.
Indicops view
Indian enterprises evaluating open-weight models should read licence terms before benchmarks. Revenue caps, territory limits and output-use restrictions can rule out a model for group companies or export-facing products, so shortlist models with permissive licences for production and keep restricted ones for research.
More in Open Source
Back to the newsroomHugging Face's chief vows to fight on for open source as Nvidia's $12.9 billion bid faces uncertainty
After Jensen Huang's exit as Nvidia CEO, Hugging Face chief Clement Delangue said the company will keep fighting for open-source AI harder than ever. Nvidia's planned $12.9 billion purchase of the platform is scheduled to close in the first half of 2027 and has drawn US Justice Department questions.
Nvidia donates OpenShell agent runtime to CNCF and previews GPU Kubernetes tooling
Nvidia has donated OpenShell, an open-source runtime that applies policy controls to autonomous AI agents, to the Cloud Native Computing Foundation. At KubeCon North America it also put new GPU Kubernetes tools and a fast inference-restore feature into public preview.
Ricursive releases Ripple-65B Preview, an Apache-licensed open-weight model built for agentic coding
Ricursive Intelligence has released Ripple-65B Preview, a 65-billion-parameter dense open-weight model pretrained for agentic coding. The company claims a top self-reported Terminal-Bench 2.1 score among US open-weight models, though the figures have not been independently reproduced and some reruns narrow the gap.