Skip to content
Neural Daily

ModelsNews

ERNIE 5.5 Flash Preview: 14B active, 1M context

🇭🇺 Olvasd el magyarul →

Baidu released ERNIE 5.5 Flash Preview, the first post-trained model on ERNIE 5.5 Flash Base. It has 280 billion parameters with 14 billion active per token and a 1-million-token context. The API is live on Qianfan, and Apache 2.0 weights are coming soon.

· 2 min read

Az ERNIE 5.5 bejelentő képe: pontokból kirajzolódó nagy 5.5 felirat kék háttéren
Image: Baidu · Press kit · Source

What Baidu released

ERNIE 5.5 Flash Preview is the first post-trained version of the previously released Flash Base. Baidu tuned it for coding, 3D modeling, knowledge work and content creation while keeping decoding fast (Baidu ERNIE).

  • 280 billion total parameters, 14 billion active per token, 1-million-token context.
  • Text-only: it cannot see images.
  • API on Baidu's Qianfan platform. Weights will be open-sourced under Apache 2.0 later.
  • An intermediate checkpoint: post-training is ongoing and Baidu says some capabilities may still fall short of the final release.

The numbers

Az ERNIE 5.5 Flash Preview benchmarktáblázata a Hy4, DeepSeek-V4.1-Flash, GLM-5.3, GPT-6.1 Sol és Claude Opus 5.5 modellekkel
Baidu · Screenshot · Source

Source: Baidu ERNIE, ERNIE 5.5 Flash Preview announcement

  • Terminal-Bench 4.0: 46.5. Hy4: 44.4, DeepSeek-V4.1-Flash: 42.3, GPT-6.1 Sol: 56.1, Claude Opus 5.5: 66.4.
  • DeepSWE v1.1: 71.2, behind DeepSeek-V4.1-Flash (77.0) and Opus 5.5 (74.2).
  • AutomationBench: 61.9, the best in the table (DeepSeek-V4.1-Flash: 60.5).
  • MCP-Atlas: 84.6, against 84.2 for Opus 5.5.
  • HLE without tools: 46.4, against 61.4 for Opus 5.5.

Scores marked with an asterisk are Baidu's own evaluations, and there is no independent check yet. Read it as a weight-class comparison: a 14B-active model competing with similar Chinese MoE models, still clearly behind closed frontier models on the hardest tasks.

What it is good for

Baidu's demos focus on long agent runs: an 11-hour coding task that produced an Unreal Engine 5.8 game in C++, an airport and a temple built through the Blender Python API, and an editable engine model in FreeCAD. These are vendor demos, but they show the target is the agent harness, not chat.

If the weights do ship under Apache 2.0, ERNIE 5.5 Flash joins the open Chinese MoE models alongside LongCat-2.5 (LongCat-2.5: a 1.6T open agent model under MIT). Until then the Qianfan API is the realistic route for European teams, so review its data handling terms first.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles