Baidu quietly drops ERNIE 5.5 Flash Base, a 280B open-weight MoE model

What's happening

Chinese search and AI giant Baidu released its new language model ERNIE 5.5 Flash Base (Chinese: 文心5.5 Flash Base) as an open-weight model on November 5, 2026. The weights are freely available on Hugging Face under baidu/ERNIE-5.5-Flash-Base, and there's a release page on ernie.baidu.com. As with many Chinese releases, the model showed up before the explanation. According to an analysis by API provider OrcaRouter, the repository was created back on October 21. Between November 2 and 3, 128 weight files in FP8 format were added, followed on November 4 by the architecture and benchmark charts. There was no press release and no launch event.

The key word is "Base": this is a base model, meaning the state right after pretraining. It hasn't yet been trained to follow instructions or chat. It continues existing text, much like your phone's autocomplete, only far more capable. It's aimed at developers and companies that want to build their own specialized models on top of it.

The specs

ERNIE 5.5 Flash Base uses a Mixture-of-Experts (MoE) architecture. The model consists of many specialized subnetworks, the "experts." For each word, only a small selection of them is consulted, a bit like a large hospital where not every doctor sees every patient. According to Baidu and OrcaRouter:

The context is the amount of text the model can take in at once. A million tokens is roughly several thick books. During training, though, the model only saw 128,000 tokens. The longer length is intended for fine-tuning and inference, and whether it holds up in practice remains to be seen.

What's new under the hood

According to Chinese tech outlet Leiphone, the design centers on efficiency. The longer the text, the more compute and VRAM it eats. Baidu cites several in-house techniques:

• Latent-Gated Sparse Attention: Attention is the mechanism the model uses to decide which earlier parts of the text matter right now. Baidu combines two variants: 37 layers with heavily compressed attention (HCA) that summarize the history, and 6 layers with MLA and sparse attention that pinpoint specific earlier positions. On top of that comes "Low-Rank Gated Attention." It reuses existing compressed intermediate values and so cuts 75 percent of the weights for the so-called gate projection. "Virtual Head Attention" is meant to improve quality while adding barely any compute at inference time.
• Multi-Router Latent MoE: Normally, a single "router" picks which experts get used. In ERNIE, several independent routers decide together. The idea is that a single misjudgment carries less weight, while compute and communication costs stay the same.
• ECHO: Many models learn to predict several tokens ahead using an extra module (Multi-Token Prediction, MTP). According to Leiphone, Baidu instead uses the model's own final layer for this. It's meant to help with planning and thus mainly improve reasoning and coding tasks. The same layer can also serve as a draft model for speculative decoding: a fast helper proposes several words, and the big model just checks them. That speeds up output. OrcaRouter, however, also mentions a 0.36B-parameter MTP head that isn't included in any of the parameter counts. The sources don't make clear exactly how these fit together.
• Training infrastructure: Baidu also reworked the distributed Muon optimizer, a method that adjusts the weights during training, and added a new attention kernel called FlashMask-4.

Benchmarks: strong, but self-reported

Baidu compares its model with three similarly sized Chinese base models: DeepSeek-V4-Flash-Base (284B, 13B active), Tencent's Hy3-preview-Base (295B, 21B active) and Xiaomi's MiMo-V2.5-Base (310B, 15B active). According to OrcaRouter, ERNIE comes out ahead on 24 of 30 tests. A few examples:

ERNIE lost on ChineseSimpleQA, HELMET-Rerank and InfiniteBench against DeepSeek, among others, and on FullStackBench and BBH with chain of thought against Hy3. Other reported scores include MMLU 93.3, GPQA 60.2 (5-shot) and LiveCodeBench 40.1 (Pass@1).

OrcaRouter warns against apples-to-oranges comparisons. A LiveCodeBench score of 40.1 looks weak next to the 91.9 of Alibaba's finished model Qwen 3.8 Flash. But one is an untuned base model, the other a fully post-trained product, measured on a newer version of the benchmark to boot. Baidu's table is internally consistent, since all models were tested with the same prompts and settings. Nobody has independently verified it yet, though.

Availability, license and deployment

The weights are released under the "ERNIE 5.5 Flash Base Model License Agreement v1.0". It's a custom license that, according to OrcaRouter, you should read in full before any commercial use. There's no API yet: the price list for Qianfan, Baidu's cloud platform, doesn't include ERNIE 5.5. For deployment, Baidu ships ready-made recipes. One is an SGLang image (lmsysorg/sglang:dev-ernie-55-flash) with tensor parallelism 4, which splits the model across four GPUs. The other runs on vLLM 0.32.0 plus the baidu/vLLM-ERNIE plugin. You access it via a completions endpoint, not a chat interface.

What you need to try it

This model won't run on a regular PC. The FP8 weights alone take up roughly 280 GB, and BF16 about twice that. Baidu's official recipe spreads the model across four data-center GPUs. Even a 4-bit quantization, a slimmed-down version that stores numbers more coarsely, would need roughly 170 GB of VRAM. There are no known adapted versions for Mac or CPU such as GGUF yet, and no Ollama entry either. In theory, workstations with lots of RAM, say 256 GB or more, could run it once the community builds such versions. There are no concrete numbers on that yet. Google Colab won't cut it. Realistically, you can only use the model on rented cloud GPUs or your own servers.

The bottom line

ERNIE 5.5 Flash Base is yet another strong open-weight base model from China. It enters a crowded class of roughly 300B parameters with 13B to 21B active. It's mainly interesting for research teams and companies that want to fine-tune their own model and are looking for a capable, efficient foundation. The many disclosed architecture details are valuable to them, too. For end users, the model offers little for now: there's no chat version, no API and no version for home hardware. The benchmarks come exclusively from Baidu. Whether a post-trained version can keep up with models like Qwen 3.8 Flash or DeepSeek remains an open question.