ERNIE 5.5 Flash Base

What it is

ERNIE 5.5 Flash Base is a language model introduced by Baidu on 5 November 2026. The weights were published on Hugging Face under the name baidu/ERNIE-5.5-Flash-Base^[1][2]. It is a base checkpoint rather than a finished assistant: Baidu released it before the mid-training and post-training stages that normally turn a pre-trained model into a usable product, and it was trained on 33.5 trillion tokens^[2]. The model works on text only^[2].

How it is built

The model is a sparse mixture-of-experts (MoE) network. In a dense network every token passes through the same parameters; in a sparse MoE, a routing step sends each token to a small number of specialised sub-networks, so the total parameter count can be much larger than the amount of computation spent on any single token.

Here the split is 280B parameters in total and 14B activated per token^[2]. The network has 43 layers: one dense layer and 42 MoE layers. Each MoE layer contains 512 routed experts, of which the top 10 are used for a given token, plus one shared expert^[2]. A shared expert is one that every token passes through, alongside the few routed experts the router picks for it.

Attention in the model is hybrid rather than uniform: 37 layers use Heavily Compressed Attention and 6 use MLA-with-DSA^[2]. Attention is the mechanism that lets a model weigh different earlier tokens when it processes each new one, and mixing two designs across the depth of the network means the model does not apply a single attention scheme everywhere.

Context length is handled in two stages. Pre-training used a 128K context, while fine-tuning and inference support up to 1 million tokens^[2]. The published weights come in two numeric formats, FP8 and BF16^[2].

Reported results

Baidu published a base-model comparison table for ERNIE 5.5 Flash Base. The reported scores are MMLU 93.3, MMLU-Pro 78.4, GPQA 60.2, HumanEval+ 82.3, LiveCodeBench 40.1 and SimpleQA 34.1^[2]. These benchmarks probe different abilities: MMLU and MMLU-Pro are multiple-choice knowledge tests, GPQA holds graduate-level science questions, HumanEval+ and LiveCodeBench measure code generation, and SimpleQA asks short factual questions.

In Baidu's table these scores lead the comparison models, which are DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base^[2]. On the long-context benchmark LongBench-v2, the table gives ERNIE 5.5 Flash Base 48.7 against 44.5 for DeepSeek-V4-Flash-Base^[2]. The same table is not a clean sweep: DeepSeek-V4-Flash-Base scores higher on ChineseSimpleQA, 75.3 versus 73.0^[2]. Since all of these numbers come from Baidu's own comparison, they are best read as the vendor's account of where its model stands rather than as an independent ranking.

Licensing and where it sits

The weights are not released under Apache 2.0. Baidu published them under the ERNIE 5.5 Flash Base Model License Agreement v1.0, a custom agreement^[2]. Anyone intending to use the model commercially therefore has to read that document rather than assume the terms of a standard open-source licence.

As a base checkpoint, the model is a starting point: it has been pre-trained but not aligned or tuned for conversation, so the usual use is further fine-tuning on a specific task or domain. The architecture is designed to keep that practical — with 14B of 280B parameters active per token, the compute needed for a forward pass is far below what the total parameter count would suggest.