AnswersPublished18 sources
What Baidu’s ERNIE-5.5-Flash-Base Release Means
Baidu released ERNIE 5.5 Flash Base on November 5, 2026: a 280B parameter mixture of experts model with 14B active per token and support for up to 1M token fine tuning and inference.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Baidu released ERNIE 5.5 Flash Base on November 5, 2026: a 280B parameter mixture of experts model with 14B active per token and support for up to 1M token fine tuning and inference.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Baidu announce on November 5, 2026, with its Hugging Face release of ERNIE-5.5-Flash-Base, and how do its 280B total and 14B active. Article summary: Baidu announced the Hugging Face release of ERNIE-5.5-Flash-Base on November 5, 2026: a downloadable, text-only mixture-of-experts base model with 280 billion total parameters and 14 billion activated per token.[3][4] Th. Topic tags: general, general web, documentation. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fak
Baidu released ERNIE-5.5-Flash-Base on Hugging Face on November 5, 2026. It is a downloadable, text-only mixture-of-experts base model: 280 billion parameters in total, with 14 billion activated for each token. Baidu says it was pretrained from scratch on 33.5 trillion tokens.1
3
6
The key distinction is that 280B describes the model’s total capacity, while 14B is the portion active per token—not the total size of the checkpoint. The release is also a starting point for adaptation, rather than a finished instruction-tuned assistant.6
What the release includes
The checkpoint is described as the state at the end of pretraining, before mid- or post-training. SGLang’s documentation presents it as a base for domain adaptation, instruction tuning and reinforcement-learning research. That makes it relevant to teams that want to customize a model, but it may require additional training to suit an assistant-style application.
The listed context limits distinguish training from later use: up to 128K tokens for training and up to 1M tokens for fine-tuning and inference. SGLang also documents deployment support for context up to 1M tokens.
The Hugging Face release includes an FP8 model, and a separate BF16 variant is listed. SGLang documents support for both.5 These are substantial model checkpoints, so parameter count alone does not describe the infrastructure needed to run them.
How its sparse architecture works
In a mixture-of-experts model, not every parameter is used for every token. ERNIE’s stated 280B total and 14B active figures reflect that sparse design.6
SGLang describes the model as having 37 heavily compressed attention layers and six full-context multi-head latent-attention layers, with sparse attention. It also lists 512 routed experts, selecting 10 alongside a shared expert in each MoE layer. These details help explain the architecture; they do not, by themselves, establish a particular end-to-end speed or cost advantage.
Baidu also reports training with its Muon optimizer and FP8 mixed precision across thousands of GPUs. Its model materials describe FlashMask-4 as an attention kernel for large head dimensions.6
For the kernel, Baidu reports up to 175% faster forward passes and 74% faster backward passes on SM100 compared with FlashAttention-4 MaskMod.4 Those are vendor-reported kernel results—not a claim that the full model runs 175% faster in a real serving workload.
How the reported benchmark results compare
Baidu’s reported results show strong scores on several knowledge and coding evaluations, including comparisons with DeepSeek-V4-Flash-Base. They are benchmark results, not independent replications.4
7
| Benchmark | ERNIE-5.5-Flash-Base | DeepSeek-V4-Flash-Base |
|---|---|---|
| MMLU-Pro | 78.4 | 69.2 |
| HumanEval+ | 82.3 | 69.5 |
| LiveCodeBench | 40.1 | 31.9 |
| ChineseSimpleQA | 73.0 | 75.3 |
On the listed MMLU five-shot evaluation, ERNIE scores 93.3, ahead of DeepSeek-V4-Flash-Base at 89.4, Hy3-preview at 87.5 and MiMo-V2.5 at 87.6.4 But the results are not a clean sweep: DeepSeek scores higher on ChineseSimpleQA and InfiniteBench in the reported comparisons.
7
14
The practical takeaway is a base model with large total capacity, sparse per-token activation, and documented long-context deployment support. Its benchmark profile is promising in several areas, but should be evaluated against the needs of a specific task—and the license—before deployment.
Serving and license considerations
SGLang lists both the FP8 and BF16 checkpoints as supported and provides deployment instructions, including ECHO speculative decoding. The Hugging Face model page identifies a custom ERNIE 5.5 Flash Base Model License Agreement v1.0. The agreement also refers to an acceptable-use policy, so availability of weights should not be taken to mean unrestricted use; review the license terms before relying on the model in a product.19
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What Baidu’s ERNIE-5.5-Flash-Base Release Means"?
Baidu released ERNIE 5.5 Flash Base on November 5, 2026: a 280B parameter mixture of experts model with 14B active per token and support for up to 1M token fine tuning and inference.