Back home

ERNIE 5.5 Flash Preview

x+web 0.96
Visit source

Intro

ERNIE 5.5 Flash Preview is the foundation model Baidu released on 11 November 2026: the first post-trained model built on the earlier ERNIE 5.5 Flash Base, officially positioned as an open-weight model designed for efficient inference. The English product page is Introducing ERNIE 5.5 Flash Preview, and the Chinese page is ERNIE 5.5 Flash Preview 发布. The company account @Baidu_Inc announced it at 2026-11-11T05:42:38Z; the model account @Baidu_ERNIE repeated the specs and said the weights will be open-sourced by 25 November.

According to the official product page, the model has 280 billion total parameters, 14 billion active per token, and a 1-million-token context, with the same limit on input and output. Officially, it is stronger at coding, 3D modeling, knowledge work and content creation, while balancing capability with inference efficiency and keeping decoding fast. @Baidu_Inc also said generation speed is above 200 tokens per second, and that it posts competitive scores on Terminal-Bench, AutomationBench and other benchmarks. Those figures come from official materials, and some comparisons may be Baidu’s own runs; this page does not treat them as independent third-party results. The API is live on Baidu’s Qianfan platform; the company later posted in the same thread that the model can be tried for free on OpenRouter. Weights are planned for release under Apache 2.0 by 25 November; they were not out at the time of this scan.

Official demos ran end-to-end in an agent harness and credit external assets and tools. For games, the page says the model works with Unreal Engine, Unity, Godot and other major engines. In one official case, it merged assets made for several Unreal Engine generations into a UE 5.8 project and wrote a third-person wuxia game, Yunxiu Xing, in C++ in a single 11-hour coding task. Because it is a text-only model and cannot see the screen, Baidu says it turned playtest problems — a character hovering, an inconsistent running direction, a misplaced camera — into measurable tests in code, then iterated with a growing test suite and playtest feedback. For 3D and design, demos include building an airport and an airliner with the Blender Python API, reconstructing Hōryū-ji, and writing editable FreeCAD models of an engine, a building and a lunar-rover concept. For knowledge work, official examples include research reports, a personal wealth-planning workbook and an original orchestral piece. For video, the page says the model can take a job from research and script through storyboard to a finished film; shorts such as The Pelican Test were drawn and composited in Python. Official cases repeatedly used tools such as Seedance 2.5 and image generators, plus third-party assets, all credited on the page.

The official page also states the limits: this is an intermediate checkpoint from post-training, some capabilities may still fall short, training is ongoing, and the official release will improve based on real-world use and feedback. Do not collapse it into the already-recorded ERNIE-5.5-Flash-Base from 5 November: Base is the base model, and Preview is its first post-trained preview.

Key links:

Explain

Why it matters

After the Flash Base release on 5 November, Baidu quickly followed with a first post-trained preview, a live Qianfan API, and a planned Apache 2.0 weight drop. For teams that want long context and stronger coding or tool use, that is a new name you can try now, with open weights promised soon. The official page spends much of its space on game engines, Blender/CAD and finished-video workflows, which is a signal that Baidu wants this model used inside an agent harness to finish projects, not only to chat.

Keep official claims and demo conditions separate. The official cases credit external assets, image and video generators, and an agent scaffold; the model itself is text-only and cannot see a game screen. Benchmark comparisons come from official materials and may include Baidu’s own runs — cite the source, and re-check the upstream page. The Preview is not the final official model.

Limits

The weights were not out at scan time; treat the open-source date as official guidance (planned by 25 November). Quotas, pricing and regional limits on Qianfan and OpenRouter should be read from those platforms at the time of use. This page does not copy specific leaderboard numbers that were not checked in the official product page or official account posts.

Usage

  1. Start with the official product page https://ernie.baidu.com/release/ernie-5-5-flash-preview/ or the Chinese page https://ernie.baidu.com/zh/release/ernie-5-5-flash-preview/ , and cross-check the announcement at https://x.com/Baidu_Inc/status/2120391732248637734 ;
  2. Call the model through the Baidu Qianfan API, or try it as the official post describes at https://openrouter.ai/baidu/ernie-5.5-flash-preview:free ; once the weights are out, use the download location given on the official page or repository ;
  3. When citing 280B/14B, the 1M-token context, 200 tokens/s, or benchmark performance, say the source is the 11 November 2026 official page or the @Baidu_Inc post, and re-check that source; when citing demos, also note the external tools and assets ;
  4. Do not describe ERNIE 5.5 Flash Preview as the finished official release, and do not merge it with ERNIE-5.5-Flash-Base; this page only restates the official pages and official accounts — re-check the upstream source before citing it.

Advertisement