Baidu Launches ERNIE 5.5 Flash Preview, Weights Open by Nov 25

What happened

The preview is the first post-trained version of ERNIE 5.5 Flash Base, released on Nov 5, keeping the same parameter scale. Baidu says in its official announcement that the preview is an intermediate checkpoint from post-training: training is not finished, so some capabilities may still fall short; post-training continues, and the final version will be improved based on the preview's real-world performance and user feedback.

The sparse mixture-of-experts model has 280 billion total parameters, activates 14 billion per token, and supports a 1-million-token context; Baidu says it generates more than 200 tokens per second. In an official demo, the model completed an 11-hour long-horizon coding task, merging assets across multiple Unreal Engine generations into a single UE 5.8 project and writing a complete xianxia game in C++; the text-only model cannot see the screen and iterates through playtest-derived code tests.

Baidu published a 20-benchmark appendix: the preview beats DeepSeek V4.1 Flash on 14 of them, and wins 2 of 8 head-to-head benchmarks against Claude Opus 5.5, GPT-6.1 Sol, Tencent Hy4 and Zhipu GLM-5.3 — AutomationBench at 61.9 and SpreadsheetBench V2 at 36.4 — while scoring 46.5 on Terminal-Bench 4.0, below Claude Opus 5.5's 63.6. All scores are vendor-reported by Baidu, with some competitor scores measured by Baidu itself, and none independently reproduced by a third party. On pricing, Qianfan uses time-of-day rates at 50% off until Nov 30: peak hours cost 1 yuan per million input tokens and 4 yuan per million output tokens, with 0.1 yuan for cache-hit input; off-peak rates are 0.5 yuan for input and 2 yuan for output.

Key facts

1. Position: First post-trained version of ERNIE 5.5 Flash Base; Baidu calls it an intermediate post-training checkpoint with training unfinished
2. Specs: 280B total parameters, 14B active per token; 1M-token context; Baidu says generation exceeds 200 tokens per second
3. Benchmarks: Beats DeepSeek V4.1 Flash on 14 of 20 Baidu-published benchmarks; wins 2 of 8 head-to-head benchmarks, all vendor-reported
4. Open source: Weights to be open-sourced under Apache 2.0 before Nov 25; no Preview weights repository on Hugging Face as of Nov 11
5. Pricing: Qianfan time-of-day rates at 50% off until Nov 30; peak input 1 yuan, output 4 yuan per million tokens