Baidu launches ERNIE 5.5 Flash Preview with 280 billion parameters

Baidu has released ERNIE 5.5 Flash Preview, a large language model with about 280 billion parameters and a context window of up to 1 million tokens.

The launch landed on November 11, 2026, just six days after its foundation model went public.

What Baidu actually shipped

The Preview is the first post-trained model in the ERNIE 5.5 Flash line. Post-training is the finishing school stage, where a raw model learns to follow instructions and behave usefully instead of just predicting text.

The headline number is 280 billion total parameters. Only 14 billion parameters activate for each token the model processes.

That comes from a Mixture-of-Experts, or MoE, architecture. A routing system sends each piece of input to a handful of specialized sub-networks, leaving the rest idle.

The practical result is a model with the knowledge base of a very large system and the running costs closer to a much smaller one. Here, the active slice works out to one-twentieth of the total.

Baidu describes the ERNIE 5.5 Flash series as using a sparse MoE design specifically for efficiency.

A million tokens of memory

At up to 1 million tokens, ERNIE 5.5 Flash Preview can theoretically take in very long documents, large codebases, or extended project histories without forgetting the beginning by the time it reaches the end.

Baidu is pitching the model for coding, 3D modeling, and content creation. The company reports competitive benchmark results across those areas, along with fast decoding speeds.

Those benchmark claims come from Baidu itself. Independent testing will determine how well they hold up once developers start pushing the model on their own workloads.

Where to get it, and when it opens up

The Preview API is live now on Baidu’s Qianfan platform, the company’s hub for developers building on its AI models.

Baidu is expected to open-source the model weights under the Apache 2.0 license by November 25, 2026.

Until that release actually happens, though, it remains a stated plan. The two-week window gives Baidu a firm public deadline to hit.

The foundation underneath

The Preview is built on ERNIE 5.5 Flash Base, which Baidu released on November 5, 2026.

The Base model was trained from scratch on 33.5 trillion tokens. Its architecture is a hybrid MoE design made up of one dense layer and 42 MoE layers. The dense layer processes everything for every token, while the MoE layers handle the specialist routing.

The Base model also supports an inference context of 1 million tokens, so the long-memory capability was baked in at the foundation level rather than bolted on later.

What this means

The research summary frames the rapid succession of releases as an effort to improve practical usability and draw community engagement.

For businesses, the sparse architecture is the part worth watching. A model that activates 14 billion parameters per token instead of all 280 billion could be cheaper and faster to run at scale.

There are caveats. The model carries the Preview label, which usually means it is not the final version. The benchmark results are self-reported, and real-world performance can differ from curated test scores.

The open-source release is also still pending. The key dates to watch are November 25, 2026, when the weights are expected to arrive, and the weeks after, when independent developers will start publishing their own results.