The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

Baidu ERNIE 5.5 Flash Preview Launches With 280B Model

By The Daily Byte ·14 Nov 2026 ·4 min read
Baidu headquarters building in Shangdi, Beijing, with the Baidu logo and an Apollo Go self-driving car outside

Baidu has unveiled ERNIE 5.5 Flash Preview, a new large artificial intelligence model with 280 billion parameters, at its annual technology event in Beijing on 12 November 2026. The launch anchors Baidu's widest AI portfolio update yet, which the company's chief executive, Robin Li, presented as a shift from building individual models to building what he calls the intelligent organisation.

ERNIE 5.5 Flash Preview is a mixture-of-experts model, a design in which only a small part of the model is active for each piece of text it processes. Baidu lists 280 billion total parameters, with only 14 billion activated per token. That sparse design is intended to deliver strong performance while keeping the computing cost of each request down.

The model also has a context window of one million tokens for inference and fine-tuning, meaning it can take in very long documents or conversations in a single request. Baidu says it was pre-trained from scratch on 33.5 trillion tokens, and that it generates text at more than 200 tokens per second.

Open weights promised by 25 November

For now, the model is available only through an application programming interface, on Baidu's Qianfan cloud platform and listed on OpenRouter. Baidu has set Qianfan pricing at RMB 1 per million input tokens and RMB 4 per million output tokens, with cached input at RMB 0.04 per million tokens.

The weights, the files that would let developers download and run the model themselves, are not yet available. Baidu says they will be fully open-sourced by 25 November 2026 under the Apache 2.0 licence. Until those files appear, the open-weight part of the announcement remains a promise rather than a delivered release.

Baidu itself has been careful about the model's status. The company describes the Preview as an intermediate checkpoint from post-training, and says that because training is not finished, some capabilities may still fall short. It plans to improve the model using feedback from users before the official release.

Benchmark claims come from Baidu alone

Baidu reported competitive scores on benchmarks including Terminal-Bench and AutomationBench, and one analysis of the launch noted a score of 46.5 on Terminal-Bench 4.0. Every figure in that comparison is vendor-reported, produced on Baidu's own test setups with judge models Baidu selected, and none of it has yet been reproduced outside the company. Independent leaderboards had not listed the model at the time of the launch.

That caution matters because benchmark tables are easy to present and hard to check. The model's headline demonstration, a third-person game written in C++ for Unreal Engine in a single 11-hour agent run, is impressive as a showcase, but it is a company demonstration rather than an independent test.

Token Exchange and the intelligent organisation

Alongside the model, Baidu introduced Token Exchange, a marketplace that connects AI model providers directly with customers. The platform is meant to help businesses discover, evaluate and buy access to model tokens from different providers, with regular spot checks on providers and real-time monitoring of API calls.

Baidu also announced upgrades to its agent products, including DuMate, a general-purpose agent for multi-step tasks, KOOKO, a document and creative workspace agent, and Famou 3.0, a self-evolving agent now spanning 25 of China's 41 industrial categories. Baidu reported that its agents now complete more than one million tasks per week, up from 84,000, and that it has built 4,300 workspaces capturing more than 20,000 organisational skills.

The company also gave an update on Apollo Go, its autonomous ride-hailing service, which it says now operates in 28 cities worldwide and has logged more than 350 million autonomous kilometres, including more than 240 million in fully driverless mode. Whether the intelligent organisation becomes the dominant way companies use AI, or remains one approach among many, will become clearer once the open weights arrive and independent testing begins.