Baidu launches ERNIE 5.5 Flash Preview, two application models and a Token Exchange
Baidu announced ERNIE 5.5 Flash Preview at Baidu World 2026 in Beijing on November 12, alongside two application models and a Token Exchange marketplace for buying model services. The new model has 280 billion total parameters with 14 billion active per token, a 1-million-token context window, and Baidu plans to open-source the weights before November 25.
📊 Scorecard: our Bullish call on $BIDU is scored from $78.14, the last close before it ran. Its first result comes after the next close.
The News
Baidu announced ERNIE 5.5 Flash Preview at Baidu World 2026 in Beijing on November 12, its latest foundation model — the large general-purpose AI model that other applications are built on. Alongside it came two application models: ERNIE Wizshot Micro-Expression, which creates video from text, image and speech, and ERNIE Light Deep Thinking, built for complex requests and tool calls at search-level speed. Baidu also launched a Token Exchange, a marketplace for buying model services.
The headline numbers are 280 billion total parameters with 14 billion active per token, and a 1-million-token context window. Parameters are the adjustable values inside a model — more of them usually means a more capable model — and “active per token” means only 14 billion of them switch on for each word the model processes, which keeps the running cost down. The context window is how much text the model can hold in mind at once; one million tokens is roughly 750,000 words, or several long books. Pricing is RMB 1 per million input tokens and RMB 4 per million output tokens, with cached input at RMB 0.04.
Baidu says it used its ERNIE Evolve technology for inference efficiency — inference is the work of running a trained model to answer questions — and reports competitive results on Terminal-Bench and AutomationBench, benchmarks that test AI models on coding and automation tasks. Those benchmark results are Baidu's own claims. The API is live on Baidu's Qianfan platform and free on OpenRouter, and Baidu plans to open-source the weights — the trained model's internal values — before November 25. Baidu flagged the Preview designation as an intermediate post-training stage, with some capabilities possibly incomplete and further improvements expected in an official version. Baidu also upgraded its DuMate, KOOKO, Baidu Yijing and Famou 3.0 agents in the same announcement.
At RMB 1 per million input tokens, ERNIE 5.5 Flash Preview is priced for volume.
Why It Matters
The price is the story as much as the model. At RMB 1 per million input tokens, ERNIE 5.5 Flash Preview is priced for volume — for developers who want to call a capable model thousands of times a day without the bill becoming the problem. That is the bet behind the Token Exchange too: if model services become cheap enough to buy and sell like a commodity, Baidu wants to be the marketplace where that trade happens.
The efficiency design is what makes the price possible. Only 14 billion of the model's 280 billion parameters are active for each token, so the model can be big in capability and small in running cost at the same time. This style of design — many specialist parts, only a few switched on per task — is now the standard way the big labs keep costs down, and Baidu is following the same playbook with its own ERNIE Evolve technology.
The open-source plan before November 25 is the other half of the strategy. Giving away the weights invites developers to build on ERNIE, test it, and find its limits — and every developer who builds on it makes Baidu's ecosystem stickier. But the Preview caveat matters: Baidu itself says some capabilities may be incomplete, so the benchmark claims should be read as the company's own until independent tests arrive.
Analyst's Take
Our call is Bullish. A capable model at commodity pricing, a marketplace already handling more than 50 million calls a day, and an open-source release within two weeks is a coherent push to make ERNIE the default cheap-and-good option for developers. The Token Exchange numbers — more than 100 model resources, over 50 million calls a day and over 700 billion tokens daily, per Baidu's figures — suggest the marketplace is already alive, not just announced.
What to watch: the weights release before November 25 and the first independent benchmarks. If the open-source weights land on time and outside tests back up Baidu's Terminal-Bench and AutomationBench claims, the model has a real shot at developer adoption beyond China. If the Preview gaps turn out to be large, the story cools fast.
Reader poll — Bullish or bearish on Baidu's AI push? Vote: Bullish, Bearish, Too early, or Watch price.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is ERNIE 5.5 Flash Preview?
Baidu's latest foundation model, announced November 12 at Baidu World 2026. “Preview” means it is an intermediate post-training stage: Baidu says some capabilities may be incomplete and further improvements are expected in an official version.
What does “14 billion active per token” mean?
The model has 280 billion parameters in total, but only 14 billion switch on for each token — each word or piece of a word — it processes. That keeps the cost of running it down while keeping the model's overall capability high.
What is the Token Exchange?
A marketplace Baidu launched for buying model services. It had already been running for a month, with more than 100 model resources, averaging over 50 million calls a day and over 700 billion tokens daily, according to Baidu's figures.
When will the model be open source?
Baidu plans to open-source the weights — the trained model's internal values — before November 25. The API is already live on Baidu's Qianfan platform and free on OpenRouter.
👁
Source: PR Newswire