On November 11 Baidu released ERNIE 5.5 Flash Preview on its ERNIE site and @Baidu_Inc. It is the first post-trained model built on ERNIE 5.5 Flash Base, whose weights came out on November 5: 280B total parameters, 14B active per token, a 1M-token context and, per Baidu, generation above 200 tokens per second. Baidu pitches it at coding, 3D modeling, knowledge work and content creation. The API is live on Baidu AI Cloud Qianfan, and the weights are due by November 25, under Apache 2.0 according to the official blog. In Baidu's own benchmark table it ranks first among the models compared on AutomationBench (61.9) and SpreadsheetBench V2 (36.4), but scores 46.5 on Terminal-Bench 4.0, well behind Claude Opus 5.5 at 63.6. Baidu says this is an intermediate post-training checkpoint.
Key Takeaways
- ✓Specs (official): 280B total, 14B active per token, 1M context, 200+ tokens/s, text-only
- ✓Vendor benchmarks: AutomationBench 61.9 (Opus 5.5: 58.0) and SpreadsheetBench V2 36.4 (Opus 5.5: 35.1), first in Baidu's comparison table on both
- ✓Weak spots too: Terminal-Bench 4.0 46.5 (Opus 5.5 63.6, GPT-6.1 Sol 56.1) and DeepSWE v1.1 71.2 (DeepSeek V4.1 Flash 77.0)
- ✓Pricing: on Qianfan, CNY 2 / 1 per million input tokens (peak / off-peak) and CNY 8 / 4 output, with a limited-time 50% discount (per NBD); OpenRouter lists $0.28 / $1.13
- ✓Weights due by November 25, under Apache 2.0 per the official blog, which is looser than the Base model's custom license; check the final release
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and industry pain point
On November 5 Baidu open-sourced the ERNIE 5.5 Flash Base checkpoint and said a post-trained model would follow. ERNIE 5.5 Flash Preview is that model. In China's open-model race, DeepSeek V4.1 Flash, GLM-5.3 and Tencent's Hy4 are all competing for the cheap, mid-sized, agent-capable slot, and Baidu needs this release to close its gap in coding and agent work.
Architecture and mechanism
It is a mixture-of-experts model with 280B total parameters, 14B active per token and a 1M-token context, and Baidu says it generates more than 200 tokens per second. Baidu describes targeted training for game engines (Unreal, Unity, Godot), Blender and FreeCAD modeling, and 300,000 finance and legal style workspaces holding 10 million reference files. It is text-only, so in Baidu's demos it turns playtest problems into code-level tests instead of looking at the screen.
Benchmarks and measured results
All figures come from Baidu's own table (vendor numbers; competitor scores marked with an asterisk were measured by Baidu). It scores 61.9 on AutomationBench and 36.4 on SpreadsheetBench V2, first among the models compared on both. On APEX-Agents 1.1 it scores 64.2, below Hy4 at 75.0. On Terminal-Bench 4.0 it scores 46.5, well behind Claude Opus 5.5 at 63.6, and on DeepSWE v1.1 it scores 71.2. There is no independent evaluation yet, and Baidu itself calls this an intermediate post-training checkpoint.
Developer guide
You can call it today on Qianfan or through OpenRouter as baidu/ernie-5.5-flash-preview, with tool calling and structured outputs. It looks best suited to spreadsheet, document and office-automation agents. For heavy terminal coding work, A/B it against your current model first. Teams that want to self-host should wait for the weights due by November 25 and check the final license.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.