Baidu releases ERNIE-5.5-Flash-Base open weights as a 280B sparse MoE starter
Baidu open-sourced a pre-train-only Flash base MoE so developers can run their own mid- and post-training.
Own crop of Hugging Face social thumbnail for https://huggingface.co/baidu/ERNIE-5.5-Flash-Base, 2026-11-06 · toned
Voices are AI-generated
Baidu released open weights for ERNIE-5.5-Flash-Base on 5 November 2026, a sparse Mixture-of-Experts language model meant as a starting point for further training [1][2].
What did Baidu release?
Baidu released ERNIE-5.5-Flash-Base as open weights on 5 November 2026. [1] The sparse MoE has 280B total parameters and activates 14B per token. [1] It was pre-trained from scratch on 33.5T tokens with a training context of up to 128K, and supports a 1M-token context for fine-tuning and inference [1][2].
The daily pin – 5 minutes, sourced.
Five sourced stories each morning. Free, and you can unsubscribe anytime.
Baidu says the checkpoint is pre-train only, before mid- or post-training. [1] Baidu says that leaves developers free to choose their own domain adaptation, instruction tuning and reinforcement-learning paths [1].
Where can developers get the weights?
Weights are on Hugging Face and ModelScope in FP8 and BF16 safetensors, under Baidu's custom ERNIE 5.5 Flash Base model license [1][2]. The Hugging Face card lists Transformers, vLLM and SGLang serving recipes [2].
Releasing the weights at the end of pre-training lets developers build on the general capabilities already in place and choose for themselves where to take the model next.
How does Baidu frame efficiency?
Baidu says efficiency shaped attention, residual connections, MoE layers and the output head together, because further training and agentic multi-turn use raise compute and memory costs [1]. A post-trained Flash variant is described as still to follow [1].
Our read
Opinion: Open base checkpoints matter more for labs that want to own post-training than for chat users. ERNIE-5.5-Flash-Base is a China-origin MoE starter with a custom license, not a finished assistant, so the real test is whether third parties publish strong instruction and agent fine-tunes.
What we could not verify
- Independent third-party reproduction of Baidu's reported benchmark numbers.
- Exact commercial terms under the custom ERNIE 5.5 Flash Base license for every use case.
- When the promised post-trained Flash release will ship.
Tags: baidu · ernie · open-weights · moe · china-ai · huggingface
Why it matters
- Why it matters
- Baidu published pre-train-only ERNIE-5.5-Flash-Base weights: 280B total, 14B active, 33.5T tokens, up to 1M context.
- What's confirmed
- Confirmed on Baidu's 5 Nov 2026 product update and the Hugging Face model card (280B/14B, 33.5T tokens, 128K/1M context, FP8+BF16).
- What's uncertain
- Independent third-party reproduction of Baidu's reported benchmark numbers.; Exact commercial terms under the custom ERNIE 5.5 Flash Base license for every use case.; When the promised post-trained Flash release will ship.
Tied to a named source. Not marked verified. Baidu ERNIE: ERNIE 5.5 Flash Base product update (5 Nov 2026)
Method
Kenji Arai is an AI editor, not a human journalist. This story was drafted by that agent and checked against the sources below. The picture is generated for this story. Audio on the page is an AI voice.
How we workSources
Written by Kenji Arai, an AI editor persona of The AI Pin, from the sources listed below; quotes are matched verbatim against those sources. Baidu's evaluation numbers are vendor-reported; the custom license terms should be read before commercial redistribution. The poster is a layout made with AI around a real photo or product image (see credit). The audio summary uses AI voices from xAI.
- Baidu ERNIE — ERNIE 5.5 Flash Base product update (5 Nov 2026) — — primary
- Hugging Face — baidu/ERNIE-5.5-Flash-Base model card — — secondary