Baidu Launches ERNIE 5.5 Flash Preview, a 280B-Parameter Model Available via API, Weights Promised by November 25
Six days after releasing its base model, Baidu has launched ERNIE 5.5 Flash Preview, the first post-trained version of its 280-billion-parameter MoE (14 billion active), already accessible via API on Qianfan. Weights are promised by November 25, with no announced license, while published benchmarks place it behind Claude Opus 5.5 on several agent tests.
Baidu announced ERNIE 5.5 Flash Preview on November 11, 2026, marking the first “finished” release of its new language model. It is a Mixture-of-Experts model with 280 billion parameters, of which only 14 billion are active per token, featuring a context window of one million tokens. Unlike the base model released six days ago, this version has been post-trained for instruction-following and agentic behavior. Its API is already live on Qianfan, Baidu’s cloud platform, and the weights are promised by November 25.
From Raw Model to Assistant
On November 5, Baidu took the opposite route of most AI labs by first releasing a raw pre-training checkpoint incapable of conversing. The Preview version represents the next step: the same architecture, designed for cost-efficient inference, but with post-training geared toward coding, agentic workflows, office productivity (spreadsheets, documents, presentations), and content generation. Baidu emphasizes speed, claiming over 200 tokens per second per user on Qianfan.
The developer remains cautious about the model’s current maturity. Its announcement notes that this is an intermediate version of post-training that may still display weaknesses stemming from incomplete training, adding that the final release of ERNIE 5.5 Flash will incorporate user feedback.
Heavy Focus on Agent Demonstrations
Most of the announcement highlights tasks allegedly completed end-to-end by the model operating within an agentic setup. The most striking showcase is a third-person action game developed in Unreal Engine 5.8 during an 11-hour coding run: the text-only model, unable to see visuals, reportedly turned every detected flaw into a measurable code-based test to iterate and advance. Baidu also showcased nearly a hundred small mobile games, 3D scenes created using Blender’s Python API (an airport with an airliner, a Japanese temple), mechanical parts designed in FreeCAD, 20- to 24-page reports, a financial modeling spreadsheet, and a 116-second animated short film drawn entirely via code.
These examples were selected by the developer and frequently rely on external assets (3D model repositories, image or video generators) detailed by Baidu. Above all, they signal the intended direction: an MoE model with a moderate active parameter count, trained to orchestrate external tools over long-running tasks.
Mixed Benchmark Results Against Claude Opus 5.5
The benchmark chart published by Baidu compares ERNIE 5.5 Flash Preview against Claude Opus 5.5, GPT-6.1 Sol, Hy4, DeepSeek V4.1 Flash, and GLM-5.3 across eight agent evaluations. The picture is more nuanced than typical promotional launches suggest. ERNIE takes the lead on AutomationBench (61.9 vs. 58.0 for Claude Opus 5.5) and SpreadsheetBench V2 (36.4 vs. 35.1). However, it trails noticeably on Terminal-Bench 4.0 (46.5 vs. 63.6 for Claude Opus 5.5 and 56.1 for GPT-6.1 Sol), DeepSWE v1.1 (71.2 vs. 77.0 for DeepSeek V4.1 Flash), APEX-Agents 1.1 (64.2 vs. 75.0 for Hy4), and Agents’ Last Exam (29.0 vs. 35.2 for DeepSeek V4.1 Flash). Baidu uses asterisks to flag figures it measured internally, which applies to several competitor results.
Promised Weights, Unknown License
Baidu bills ERNIE 5.5 Flash as an open-weights model and promises to release them by November 25. However, the reviewed documentation says nothing about the license that will govern them. This is a critical point for European users: the base model’s license excludes the European Union, the United Kingdom, and South Korea, and prohibits using its outputs to train or improve competing models. If these restrictions carry over, the Preview release will not be legally usable in France, even after the weights go live.
Our take
By rolling out the base checkpoint first followed swiftly by a post-trained preview, Baidu is rapidly catching up with the cadence of its domestic rivals. The technical trade-offs make sense: 14 billion active parameters, a one-million-token context window, and high inference speeds are tailored for persistent agents, where cost per run matters just as much as benchmark scores. Yet the published benchmark table remains candid on one key reality: on terminal and software engineering benchmarks, Claude Opus 5.5, GPT-6.1 Sol, and DeepSeek V4.1 Flash remain ahead, and several competitor scores were measured by Baidu itself. Independent benchmarking will be required to properly gauge its standing. For Europe, the more pressing issue lies elsewhere: until Baidu clarifies the model weights’ license, there is no guarantee this “open” release will be accessible to European companies.
Sources
- Baidu — ERNIE 5.5 Flash Preview, launch post (November 11, 2026, in Chinese) Primary source
- Baidu Inc. — announcement on X (November 11, 2026) Primary source
- Baidu ERNIE — announcement on X with benchmark chart image (November 11, 2026) Primary source
- Baidu — ERNIE 5.5 Flash Base, technical post (November 5, 2026) Primary source
- Baidu — ERNIE 5.5 Flash Base model license (Hugging Face) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
