Baidu · 2026-11-11 · major
ERNIE 5.5 Flash Preview — Baidu's 280B MoE for agentic coding
In short
ERNIE 5.5 Flash Preview is Baidu's first post-trained model on ERNIE 5.5 Flash Base: 280B total and 14B active parameters, 1M-token context. It is on Qianfan and free on OpenRouter; Apache-2.0 weights are promised.
Baidu · 2026-11-11 · 3 sources checked
Baidu's 280B mixture-of-experts model with 14B active parameters, built for long agentic coding and office work.
Key specs
| Parameters | 280B |
|---|---|
| Active params | 14B |
Quick facts
| Maker | Baidu |
|---|---|
| Architecture | MoE, 280B total / 14B active |
| Context window | 1M tokens |
| Availability | Baidu Qianfan API + OpenRouter (free) |
| Weights | Apache-2.0, announced, no date |
| Status | Preview, an intermediate post-training checkpoint |
Benchmarks
| ERNIE 5.5 Flash Preview | 61.9% | |
|---|---|---|
| Claude Opus 5.5 | 58% | |
| DeepSeek V4.1 Flash 1017 | 57.2% | |
| Hunyuan Hy4 | 55.2% | |
| GPT-6.1 Sol | 49.3% |
| Claude Opus 5.5 | 63.6% | |
|---|---|---|
| GPT-6.1 Sol | 56.1% | |
| ERNIE 5.5 Flash Preview | 46.5% | |
| Hunyuan Hy4 | 42.4% | |
| DeepSeek V4.1 Flash 1017 | 40.9% |
What is it?
ERNIE 5.5 Flash Preview is the first post-trained model built on ERNIE 5.5 Flash Base, the 280B base checkpoint Baidu released on November 5. Baidu calls the preview an intermediate checkpoint: post-training is still running, and the official release will include improvements.
How does it work?
The model is a sparse mixture-of-experts: only 14B of its 280B parameters run for each token, which keeps decoding fast for a model this size. It has a 1M-token context window. Baidu shows it on long agent tasks, such as an 11-hour coding run that built a third-person game in Unreal Engine 5.8, Blender scenes and FreeCAD models.
Why does it matter?
Baidu's own table puts the preview ahead of Claude Opus 5.5 and GPT-6.1 Sol on AutomationBench and SpreadsheetBench V2, but well behind both on Terminal-Bench 4.0, and many competitor scores are Baidu's own runs. With a free OpenRouter endpoint and Apache-2.0 weights promised, developers get another cheap Chinese option for agent work.
Who is it for?
developers building coding and office agents
Frequently asked questions
- Are ERNIE 5.5 Flash Preview weights open source?
- Not yet. Baidu says the ERNIE 5.5 Flash Preview weights will be released under Apache 2.0 soon, without a date. The base checkpoint, ERNIE 5.5 Flash Base, is already on Hugging Face in FP8 and BF16 under Baidu's own ERNIE 5.5 Flash Base Model License.
- Where can I use ERNIE 5.5 Flash Preview?
- Baidu offers ERNIE 5.5 Flash Preview through its Qianfan API platform and as a free model on OpenRouter under the id baidu/ernie-5.5-flash-preview:free. Baidu's announcement does not list API prices, so the free OpenRouter route is the simplest way to test the model.
- How does ERNIE 5.5 Flash Preview compare with Claude Opus 5.5 and GPT-6.1 Sol?
- In Baidu's table, ERNIE 5.5 Flash Preview scores 61.9 on AutomationBench against 58.0 for Claude Opus 5.5 and 49.3 for GPT-6.1 Sol. On Terminal-Bench 4.0 it scores 46.5, behind GPT-6.1 Sol at 56.1 and Claude Opus 5.5 at 63.6. Baidu ran many competitor scores itself.
- What are the limits of the ERNIE 5.5 Flash Preview release?
- Baidu describes ERNIE 5.5 Flash Preview as an intermediate checkpoint from post-training. Baidu says some capabilities may still fall short, that training continues, and that the official ERNIE 5.5 Flash release will improve on it. The model is text-only; Baidu used a Claude Opus 5.5 tool for image understanding in BlenderWorld.
Try it
https://openrouter.ai/baidu/ernie-5.5-flash-preview:free