Baidu ERNIE 5.5 Flash Base Explained: 280B Open Model, 1M Context, and the License Catch (World Tech)

Illustrative image: Baidu's headquarters in Shangdi, Beijing. No freely licensed image of the ERNIE 5.5 Flash Base model itself exists. Image: Wikimedia Commons.

TL;DR — Baidu ERNIE 5.5 Flash Base explained in one line: on 5 November 2026 Baidu released the downloadable weights of a 280-billion-parameter AI model that activates only 14 billion parameters per token and handles up to 1 million tokens of context — but it is a raw, untuned base model under a custom license that excludes the EU, UK and South Korea.

📌 KEY HIGHLIGHTS

• Released today: Baidu's ERNIE team published ERNIE-5.5-Flash-Base on 5 November 2026, with weights on Hugging Face in FP8 and BF16 formats.
• The size: a sparse Mixture-of-Experts model with 280 billion total parameters, of which only 14 billion are activated per token — big brain, small bill per word.
• The training: pre-trained from scratch on 33.5 trillion tokens, with a 128K training context that extends to 1 million tokens for fine-tuning and inference.
• The catch: it is a base model, not a chatbot — no instruction tuning, no chat alignment, no refusal policy. And its custom license bars use in the EU, UK and South Korea.
• The claim: Baidu says it ranked first in 24 of 30 benchmarks against DeepSeek-V4-Flash-Base, Tencent Hy3-preview-Base and Xiaomi MiMo-V2.5-Base — Baidu's own tests, not independently reproduced.

Baidu ERNIE 5.5 Flash Base explained: what happened?

Baidu has open-sourced a giant AI model again. On 5 November 2026, its ERNIE team released ERNIE-5.5-Flash-Base, the first Baidu model with over 100 billion parameters to get downloadable weights in about 16 months — since the ERNIE 4.5 series in June 2025. The ERNIE 5.0 and 5.1 models in between were never open-sourced, so this is Baidu's return to the open-weights race after a long quiet stretch.

The numbers are the story. The model has 280 billion parameters in total, but it is a sparse Mixture-of-Experts (MoE) design: for each token it generates, only 14 billion parameters actually switch on. Think of it as a hospital with 280 specialists on staff, where only the 14 relevant ones are called into the room for each patient. That is why Baidu calls it "Flash" — the per-token cost stays low while the model's total knowledge stays large. It was pre-trained from scratch on 33.5 trillion tokens, using the Muon optimiser and FP8 mixed precision, and it supports a context window of up to 1 million tokens — roughly 700,000 words, or several whole books at once.

Under the hood, the architecture has 43 Transformer layers, a hidden size of 4,096 and 64 attention heads. Thirty-seven of the layers use a trick Baidu calls Heavily Compressed Attention: every 128 tokens of history get squeezed into a single cached entry, while the remaining six layers keep fine-grained access to the original text. That is how the model keeps long-context costs down without forgetting what it read. Each layer (except the first) routes tokens through 512 experts, activating 10 of them plus one shared expert.

What a base model is — and what it is not

The word "Base" in the name matters more than the parameter count. This checkpoint is the result of pre-training only — the stage where a model learns to predict the next word from raw text. It has had no mid-training and no post-training: no instruction tuning, no chat alignment, no refusal policy, no tool-use format. Download it and ask it a question, and you will get raw text continuation, not a helpful assistant.

Attribute	ERNIE 5.5 Flash Base
Total parameters	280 billion (sparse MoE)
Activated per token	14 billion
Pre-training data	33.5 trillion tokens
Context length	128K trained; up to 1M tokens for fine-tuning/inference
Architecture	43 layers, 512 routed experts (10 active) + 1 shared expert
Weight formats	FP8 (~280GB) and BF16 (~560GB) on Hugging Face
Serving	SGLang and vLLM; NVIDIA Megatron Bridge SFT recipe
License	Custom ERNIE 5.5 Flash Base License v1.0 — excludes EU, UK, South Korea

That is by design. Baidu is handing researchers and developers a fixed, common starting point: a checkpoint they can fine-tune for a domain, teach to follow instructions, or train with reinforcement learning — without repeating a 33.5-trillion-token pre-training run themselves. It also gives the research community a fixed baseline for controlled comparisons of training methods. Baidu says a post-trained ERNIE 5.5 Flash — the chat-ready version — will follow this base release.

The entrance of Baidu's headquarters campus in Shangdi, Beijing. Image: Wikimedia Commons.

Why it matters for World Tech readers

Three things make this release worth watching. First, the timing: it lands one day after Tencent unveiled Hy4, its 770-billion-parameter model with vision and computer use — covered on this blog yesterday. China's two biggest AI labs are now releasing flagship open-weight models within 24 hours of each other, and both are aimed at the same audience: developers building agents. The Hy4 launch promised weights later; Baidu shipped weights on day one, which is a real difference for researchers who want to run the model themselves.

Second, the efficiency story. A 1-million-token context in an open-weight base model is genuinely useful for agent research: an agent that can hold an entire codebase, a long legal file or a multi-hour meeting transcript in context without re-reading it. The compressed-attention design is Baidu's answer to the same problem every lab is fighting — long context is powerful, but the memory cost grows with every token. Whether the 24-of-30 benchmark wins hold up under independent testing is an open question: Baidu ran the comparisons itself, on its own harness, and the scores have not been independently reproduced.

Third, the license is the quiet twist. ERNIE 4.5 shipped under the permissive Apache 2.0 license; ERNIE 5.5 Flash Base switches to a custom license that does not authorise use, modification or distribution in the European Union, the United Kingdom or South Korea, and requires separate written authorisation from Baidu for commercial use by entities with more than $50 million in revenue over the past 12 months. For a model marketed as open, that is a meaningful restriction — and it is exactly the kind of term that matters more than the parameter count for anyone planning to build on it. The 753-billion-parameter Hydrogen-1.0 science model showed the same week that open-weight releases are getting bigger; the ERNIE release shows their licenses are getting more careful.

Should you care?

✅ Yes, if you build or study AI models. A 280B base checkpoint with 1M-token context, downloadable today and serving-ready on vLLM and SGLang, is a serious research asset — especially for fine-tuning and RL experiments. ⏳ Don't treat it as a chatbot or as fully open. It has no instruction tuning, needs roughly 280GB (FP8) to 560GB (BF16) of weights before you count memory for the context, and its license excludes the EU, UK and South Korea. Read the license before you build on it — and wait for the post-trained version if you want an assistant, not a research starting point.

What happens next: watch for Baidu's post-trained ERNIE 5.5 Flash, for independent benchmark reproductions of the 24-of-30 claim, and for whether other Chinese labs follow Baidu's lead in tightening open-weight licenses.

FAQs

What is Baidu ERNIE 5.5 Flash Base?
It is a sparse Mixture-of-Experts language model released by Baidu on 5 November 2026, with 280 billion total parameters and 14 billion activated per token. It is a base model — the result of pre-training only — released with downloadable weights for researchers and developers to fine-tune.

Is ERNIE 5.5 Flash Base really open source?
Not in the usual sense. The weights are downloadable, but under Baidu's custom ERNIE 5.5 Flash Base Model License Agreement v1.0, not an open-source license like Apache 2.0. Use, modification and distribution are not authorised in the European Union, the United Kingdom or South Korea, and large commercial users need separate written authorisation from Baidu.

How big is ERNIE 5.5 Flash Base?
280 billion parameters in total, with 14 billion activated per token. The FP8 weights need roughly 280GB of storage and the BF16 version roughly 560GB, before accounting for memory for the context window and runtime overhead. It was pre-trained on 33.5 trillion tokens and supports up to 1 million tokens of context.

Can I use ERNIE 5.5 Flash Base for commercial projects?
It depends where you are and how big you are. The license excludes the EU, UK and South Korea entirely, and entities with more than $50 million in revenue over the past 12 months need Baidu's separate written authorisation for commercial use. Smaller users elsewhere should still read the full license text on the model card before deploying it.

TAKEAWAY: Baidu's biggest open release in 16 months is a research tool, not a product — a 280B starting point for fine-tuning, with a 1M-token context and a license that quietly draws borders around who can use it. The weights are the headline; the license is the fine print that decides who benefits.

Follow World Tech on TechByCountry for tomorrow's update.

📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.

Sources

• Baidu ERNIE team — ERNIE 5.5 Flash Base release page (5 November 2026)
• Unite.AI — Baidu Releases ERNIE 5.5 Flash Base for Further Training (5 November 2026)
• ChainCatcher — Baidu open-sources the 280-billion-parameter ERNIE 5.5 Flash model after a 16-month gap (5 November 2026)
• AlphaSignal — Baidu Drops ERNIE 5.5 Flash, a Raw 280B Open Checkpoint Beating DeepSeek (5 November 2026)

Related reading on TechByCountry: Tencent Hunyuan Hy4 AI model: vision, computer use and the open-weights wait; Deep Principle Hydrogen-1.0: the 753B open-weight AI for chemistry and materials; DeepSeek, Moonshot and GPT-6 Astra distillation: Kimi K3.1 Pro, DeepSeek V4.1-Flash and the copying row.