# ERNIE 5.5 Flash Preview

> Baidu's first post-trained model on ERNIE 5.5 Flash Base, released November 11, 2026: a 280B-total / 14B-active MoE with a 1M-token context, tuned for agentic coding, 3D modeling, knowledge work and content creation.

- Maker: Baidu
- Line: ERNIE (open weights)
- Canonical page: https://ai-tldr.dev/models/ernie-5-5-flash-preview/
- Updated: 2026-11-11

## In short

ERNIE 5.5 Flash Preview is a language model from Baidu, released on November 11, 2026. It has a 1M-token context window. You can use it through Baidu Qianfan and OpenRouter.

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 1M | [1] |
| Licence | Proprietary API (Apache-2.0 weights announced) · API only | [1] |
| Released | 11 November 2026 | [2] |
| Parameters | 280B total · 14B active | [1] |

## Sources

1. [Baidu Qianfan — ERNIE 5.5 Flash Preview](https://console.bce.baidu.com/qianfan/modelcenter/model/buildIn/detail/am-abe18d298d70)
2. [ERNIE blog — Introducing ERNIE 5.5 Flash Preview](https://ernie.baidu.com/release/ernie-5-5-flash-preview/)

## Overview

ERNIE 5.5 Flash Preview is a language model from Baidu, released on November 11, 2026. It is the first post-trained model built on ERNIE 5.5 Flash Base, the open pre-trained checkpoint Baidu published on November 5, 2026, and keeps that model's size: 280B total parameters with 14B active per token and a 1M-token context. It is a text-only model.

Baidu describes it as an intermediate checkpoint from post-training rather than the finished ERNIE 5.5 Flash: it says the model is already competitive with models of its size, but that some capabilities may still fall short because training is not finished, and that the official release will follow. At launch the API is served on Baidu's Qianfan platform and on OpenRouter; Baidu says the weights will be open-sourced under Apache 2.0 soon, so the open-weights flag on this page changes only when they are out.

Baidu's launch post centers on four areas. For game development it trained on domain data with a dedicated evaluation framework and added support for Unreal Engine, Unity and Godot; one showcase is a third-person Unreal Engine 5.8 game written entirely in C++ in a single 11-hour coding task. For 3D modeling and design it trained on Blender and CAD work, with demos of parametric scenes built through the Blender Python API and engine and building models scripted in FreeCAD. For knowledge work it built 300,000 finance and legal workspaces holding 10 million reference files, and shows the model producing slide decks, spreadsheets and PDF reports. For video creation it built more than 150,000 creation environments covering script, storyboard, assets and voice-over.

In Baidu's comparison table, scored at each model's highest reasoning setting, it reaches 46.5 on Terminal-Bench 4.0, 71.2 on DeepSWE v1.1, 61.9 on AutomationBench, 84.6 on MCP-Atlas, 87.4 on BrowseComp and 46.4 on the text-only subset of HLE. It leads that table on AutomationBench, SpreadsheetBench V2 and MCP-Atlas, while Claude Opus 5.5 and GPT-6.1 Sol are ahead on Terminal-Bench 4.0, Terminal-Bench-Science and HLE, and Hunyuan Hy4 is ahead on APEX-Agents 1.1 and DRACO.

## Benchmarks

- BrowseComp: 87.4% ([source](https://ernie.baidu.com/release/ernie-5-5-flash-preview/))
- MCP-Atlas: 84.6% ([source](https://ernie.baidu.com/release/ernie-5-5-flash-preview/))
- Toolathlon-Verified: 75% ([source](https://ernie.baidu.com/release/ernie-5-5-flash-preview/))
- DeepSWE v1.1: 71.2% ([source](https://ernie.baidu.com/release/ernie-5-5-flash-preview/))
- OfficeQA Pro V2: 67.9% ([source](https://ernie.baidu.com/release/ernie-5-5-flash-preview/))
- AutomationBench: 61.9% ([source](https://ernie.baidu.com/release/ernie-5-5-flash-preview/))
- Terminal-Bench 4.0: 46.5% ([source](https://ernie.baidu.com/release/ernie-5-5-flash-preview/))
- HLE (text-only, no tools): 46.4% ([source](https://ernie.baidu.com/release/ernie-5-5-flash-preview/))
- SpreadsheetBench V2: 36.4% ([source](https://ernie.baidu.com/release/ernie-5-5-flash-preview/))

## Strengths

- 280B-total / 14B-active MoE with a 1M-token context, post-trained from an open base model
- Trained for long-horizon coding, including game development in Unreal Engine, Unity and Godot
- Blender and FreeCAD scripting for parametric 3D modeling and design
- 61.9 on AutomationBench and 84.6 on MCP-Atlas, the top scores in Baidu's comparison table

## Best for

- Reach for it for agentic coding and game-development tasks driven through a coding-agent harness.
- Reach for it to script 3D scenes and CAD parts through the Blender Python API or FreeCAD.
- Reach for it to turn a folder of reference files into a spreadsheet, slide deck or report.
- Look elsewhere if you need image input or downloadable weights today: it is text-only and the Apache 2.0 weights were not out at release.

## How to access

- [Baidu Qianfan](https://console.bce.baidu.com/qianfan/modelcenter/model/buildIn/detail/am-abe18d298d70)
- [OpenRouter](https://openrouter.ai/baidu/ernie-5.5-flash-preview:free) — model id `baidu/ernie-5.5-flash-preview:free`

## ERNIE (open weights) versions

- ERNIE 5.5 Flash Preview (11 November 2026): First post-trained model on the 5.5 Flash base; API-only at release, Apache 2.0 weights announced. — this page
- [ERNIE 5.5 Flash Base](https://ai-tldr.dev/models/ernie-5-5-flash-base/) (5 November 2026): Open-weight pre-trained base checkpoint, 280B total / 14B active.

## FAQ

### What is ERNIE 5.5 Flash Preview?

ERNIE 5.5 Flash Preview is a language model from Baidu, released on November 11, 2026. It is the first post-trained model built on ERNIE 5.5 Flash Base: a Mixture-of-Experts model with 280B total parameters, 14B active per token and a 1M-token context, aimed at coding, 3D modeling, knowledge work and content creation.

### How is it different from ERNIE 5.5 Flash Base?

ERNIE 5.5 Flash Base, released November 5, 2026, is the raw checkpoint at the end of pre-training with no instruction tuning. ERNIE 5.5 Flash Preview is that base after post-training, so it follows instructions and works inside agent harnesses. Baidu calls the Preview an intermediate post-training checkpoint and says the official ERNIE 5.5 Flash will follow.

### Are the ERNIE 5.5 Flash Preview weights open?

Not at release. Baidu's launch post says the weights will be open-sourced under Apache 2.0 soon; on November 11, 2026 the model was available only through the API on Baidu Qianfan and OpenRouter.

### How does ERNIE 5.5 Flash Preview compare with Claude Opus 5.5 and GPT-6.1 Sol?

In Baidu's table it is ahead of both on AutomationBench (61.9 vs 58.0 and 49.3) and SpreadsheetBench V2 (36.4 vs 35.1 and 32.0), and level with Claude Opus 5.5 on MCP-Atlas (84.6 vs 84.2). Both are ahead on Terminal-Bench 4.0, Terminal-Bench-Science 0.1 and HLE. Many competitor figures in that table come from Baidu's own evaluations.

### Can ERNIE 5.5 Flash Preview see images?

No. Baidu describes it as a text-only model. In its BlenderWorld evaluation, image understanding was supplied through a separate tool.

## Links

- [ERNIE blog — Introducing ERNIE 5.5 Flash Preview](https://ernie.baidu.com/release/ernie-5-5-flash-preview/)
- [Baidu Qianfan — ERNIE 5.5 Flash Preview](https://console.bce.baidu.com/qianfan/modelcenter/model/buildIn/detail/am-abe18d298d70)
- [OpenRouter — baidu/ernie-5.5-flash-preview](https://openrouter.ai/baidu/ernie-5.5-flash-preview:free)
- [ERNIE blog — Introducing ERNIE 5.5 Flash Base](https://ernie.baidu.com/release/ernie-5-5-flash-base/)
- [Baidu ERNIE](https://ernie.baidu.com/)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/ernie-5-5-flash-preview/