Baidu released ERNIE 5.5 Flash Base on 5 November 2026 — a sparse mixture-of-experts base model with 280 billion total parameters and 14 billion active per token, trained from scratch on 33.5 trillion tokens. It has a 128K training context and a 1-million-token context for fine-tuning and inference.
The weights can be downloaded from Hugging Face in FP8 and BF16 (about 286 GB for the FP8 weights) and served with SGLang and vLLM. This is the raw checkpoint at the end of pre-training, before any instruction tuning: it cannot follow instructions or hold a conversation out of the box, and a chat-ready ERNIE 5.5 Flash is still to come.
The licence is Baidu's own agreement — not Apache or MIT — and its territory clause excludes the EU, the UK and South Korea.
Key facts
| Item | Details |
|---|---|
| Release | 5 November 2026 |
| Model | Sparse mixture-of-experts base model |
| Parameters | 280 billion total; 14 billion active per token |
| Training | From scratch on 33.5 trillion tokens |
| Context | 128K training; 1 million tokens for fine-tuning and inference |
| Weights | Hugging Face, FP8 and BF16 (about 286 GB for FP8) |
| Serving | SGLang and vLLM |
| Status | Raw pre-training checkpoint — no instruction tuning yet; chat-ready version still to come |
| Licence | Baidu's own agreement (not Apache or MIT); territory clause excludes EU, UK, South Korea |
What this means
This is a base model, not a chatbot — and that distinction matters. Out of the box it cannot follow instructions or hold a conversation; it is a starting point for teams that want to do their own fine-tuning. For researchers and labs with the hardware to run 280-billion-parameter weights, it is a large, openly downloadable checkpoint to build on.
Two cautions. First, the licence: Baidu's own agreement is not Apache or MIT, and the territory clause excludes the EU, the UK and South Korea — read it before using the weights commercially. Second, Baidu's own benchmark claims (MMLU 93.3, HumanEval+ 82.3, and beating DeepSeek V4 Flash Base on 26 of 30 tests in its own table) have no independent verification yet — treat them as the vendor's claims, not as settled results.
A note on verification
The release details are from the Hugging Face model card (baidu/ERNIE-5.5-Flash-Base), gigazine (6 November 2026), Deyron Labs, frontierdot.com, forkalt.com and eesel.ai, following a Quora lead by AI Tech Tools. The lead is credited in the source line; all content here is written in our own words. The 5 November date is just outside this pass's two-week window; the release is included because it is a significant open-weights release.
FAQs
What is ERNIE 5.5 Flash Base?
A sparse mixture-of-experts base model from Baidu, released 5 November 2026, with 280 billion total parameters and 14 billion active per token, trained from scratch on 33.5 trillion tokens.
Can it chat out of the box?
No. It is the raw checkpoint at the end of pre-training, before instruction tuning — it cannot follow instructions or hold a conversation. A chat-ready ERNIE 5.5 Flash is still to come.
Where can the weights be downloaded?
From Hugging Face in FP8 and BF16 (about 286 GB for the FP8 weights). They can be served with SGLang and vLLM.
What licence does it use?
Baidu's own agreement — not Apache or MIT. Its territory clause excludes the EU, the UK and South Korea.
Source: Baidu release, 5 November 2026 (Hugging Face model card baidu/ERNIE-5.5-Flash-Base; gigazine; Deyron Labs; frontierdot.com; forkalt.com; eesel.ai); Quora lead: AI Tech Tools (https://www.quora.com/What-is-Baidus-Ernie-large-language-model-LLM-used-for/answer/AI-Tech-Tools-1?prompt_topic_bio=1). Last checked: 21 November 2026.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!