Baidu drops 280B ERNIE 5.5 Flash Preview for free on OpenRouter
Baidu released a chat-ready 280B model you can try for free today, promising Apache 2.0 open weights before November 25.
Models · Source: Baidu ERNIE
What happened
Baidu just released ERNIE 5.5 Flash Preview. It is a chat-ready version of the 280-billion-parameter base model they dropped last week. The model activates 14 billion parameters per token and handles a one-million-token context window. You can test it for free right now on OpenRouter.
The model is an intermediate post-training build. Baidu claims it excels at long-horizon tasks like coding, 3D modeling, and content creation. In one demo, the model reportedly spent 11 hours writing C++ code to build an entire third-person martial arts game in Unreal Engine 5.8. It also generated 20-page research reports and wrote Python scripts to design 3D engines in FreeCAD. The base model was trained from scratch on 33.5 trillion tokens. For this preview, Baidu trained it on 300,000 knowledge-work workspaces containing 10 million reference files.
Baidu is making a serious push for developer adoption. The company promised to release the model weights under a permissive Apache 2.0 license by November 25. This contrasts with the base model, which shipped on November 5 under a custom, restrictive license. Baidu also launched a Token Exchange marketplace to connect AI model providers with customers, processing 50 million daily calls across 100 model resources.
Key facts
- 280 billion — Total parameters in ERNIE 5.5 Flash Preview
- 14 billion — Active parameters per token
- 1 million — Maximum token context window
- November 25 — Deadline for Baidu to release open weights
- RMB 0.04 — Cost per million cached input tokens on Qianfan
Why it matters
Builders get another massive open-weight checkpoint to play with. If Baidu honors the Apache 2.0 promise, you can run and commercialize a 280B Chinese model without legal friction. The one-million-token context and strong reported coding capabilities make this a serious candidate for agentic workflows. You can prototype against the free OpenRouter endpoint today and swap to self-hosting once the weights drop.
The launch of the Baidu Token Exchange signals a shift in how model compute is traded. Baidu CEO Robin Li calls this the era of the intelligent organization. By creating a clearinghouse that verifies model fingerprints and conducts real-time monitoring, Baidu is turning model inference into a commodity market. This forces smaller API providers to compete purely on price and reliability. It also gives enterprises a transparent way to buy trusted tokens without vendor lock-in.
For builders
Free prototyping on OpenRouter today
You can test the model immediately without spending a dime. Once the weights drop, developers who need data privacy can host it themselves using vLLM or SGLang. Proprietary model providers lose leverage when 280B models become free to download.
Cheap API pricing on Qianfan
Baidu is aggressively undercutting competitors on its own cloud. The API costs just RMB 1 per million input tokens and RMB 4 per million output tokens. Builders relying on heavy context caching pay only RMB 0.04 per million tokens, saving massive costs on agentic loops.
Long context agentic workflows
The one-million-token window allows you to dump entire codebases or massive document libraries into the prompt. Startups building autonomous coding agents or complex financial analysis tools win big here. Companies selling expensive RAG pipelines lose out as native context windows grow.
My take
I love seeing Baidu sweat. Promising Apache 2.0 for a 280B model shows they know they have to give away the farm to win developer mindshare. If the weights actually drop on November 25, this becomes a massive gift for anyone building local agents.
Original reporting: Baidu ERNIE. This is my rewrite and opinion.