Original title: Hands-on test of Baidu ERNIE 5.5 Flash Preview: Aesthetics remain strong, programming takes the lead
According to Leiphone, Baidu's ERNIE 5.5 Flash Preview is now live on the Qianfan platform, with both the web interface and API available simultaneously. The official statement says the weights will be open-sourced under the Apache 2.0 license by 11 25. Just a week ago, Baidu open-sourced the base model, ERNIE 5.5 Flash Base. TMTPost reports that the new model uses a sparse MoE architecture, with official figures stating 2800 billion total parameters and 140 billion active parameters. It supports a context window of one million tokens and primarily targets programming, 3D modeling, knowledge work, and content creation. API pricing is 2 yuan per million input tokens during peak hours and 1 yuan during off-peak hours; output is 8 yuan per million tokens during peak hours and 4 yuan during off-peak hours, currently with a limited-time 50% discount. The official statement claims it ranks first on the AutomationBench and SpreadsheetBench V2 benchmarks.
Why it matters ERNIE released an open-source base model and a new model within a week, focusing on speed and cost for programming and agentic tasks. Once the weights are open-sourced, local deployment will be possible.
Credibility: Fairly credible · Sources include Leiphone and 2 others · Event time 11-11 (Beijing Time)
Reasoning
- Both Leiphone and TMTPost provided hands-on reports, detailing the launch channels, pricing, and open-source plans
Usage
- Call via the web interface or API on Baidu's Qianfan platform; after open-sourcing, it can be deployed using SGLang or vLLM
Risks
- Benchmark results are comparisons provided by Baidu, some of which use vendor self-assessment metrics
- It is still a Preview version, and the weights have not yet been released