What happened
Baidu released a preview of its ERNIE 5.5 Flash language model, a 280 billion‑parameter mixture‑of‑experts system that activates 14 billion parameters per token and supports a 1 million‑token context window. The preview is accessible for free through the OpenRouter API (identifier baidu/ernie-5.5-flash-preview:free) as long as the preview period lasts. The model runs on Baidu’s Qianfan cloud platform, and the API endpoint does not support OpenRouter’s EU regional routing. Baidu promises to publish the model’s weights under an Apache 2.0 license by 25 November 2026, though the repository was not yet available as of 13 November. The preview adds post‑training to the previously released ERNIE 5.5 Flash Base (released 5 November) and includes benchmarks showing competitive performance on tool‑heavy tasks versus Chinese rivals, though it lags behind closed‑source models on pure knowledge tasks.
On 11 November 2026 Baidu announced the ERNIE 5.5 Flash Preview, a mixture‑of‑experts (MoE) language model with 280 billion total parameters but only 14 billion active per token. The model supports a 1 million‑token context window and is limited to text‑only input and output.
The preview is offered for free via OpenRouter’s API (baidu/ernie-5.5-flash-preview:free) as long as the preview period lasts. Users must have an OpenRouter account and API key; requests are processed by Baidu’s Qianfan cloud platform. The endpoint does not support OpenRouter’s EU regional routing, raising GDPR considerations for prompts containing personal data of European citizens.
Baidu describes the preview as an intermediate checkpoint that adds post‑training to the ERNIE 5.5 Flash Base released on 5 November. The Base model’s weights are under a restrictive license (non‑commercial free, commercial free up to $50 million revenue, with geographic exclusions). Baidu promises to release the preview weights under an Apache 2.0 license by 25 November 2026, though the repository was not yet live as of 13 November.
Benchmark tables published by Baidu compare the preview against DeepSeek V4.1 Flash 1017, Tencent’s Hy4, OpenAI’s GPT‑6.1 Sol, and Anthropic’s Claude Opus 5.5. ERNIE 5.5 Flash Preview leads Chinese rivals on tool‑heavy tasks (e.g., Terminal‑Bench 4.0, AutomationBench) but trails closed‑source models on pure knowledge and browsing benchmarks.
Demo videos highlight the model’s ability to generate an entire Unreal Engine 5.8 game in C++ over an 11‑hour run and to produce nearly 100 smartphone mini‑games from design documents, showcasing Baidu’s focus on long‑duration agent workflows in creative tools such as Blender, Unity, and Godot.
Source details: pasqualepillitteri.it
Why it matters
The preview gives developers worldwide immediate, cost‑free access to a large‑scale, long‑context language model that can run agent‑style workloads for hours, a segment where price and context length are critical. If Baidu follows through on the Apache 2.0 release, the model would become one of the few openly licensed Chinese‑origin LLMs, removing geographic and revenue‑threshold restrictions that currently limit the Base model’s use in Europe, the UK, and South Korea. This could accelerate adoption of Chinese‑developed AI in global developer ecosystems, increase competition for open‑source models, and provide a new option for long‑context applications such as code generation, 3D modeling, and multi‑step reasoning. However, the model’s text‑only capability, lack of EU‑regional routing, and pending license verification introduce compliance and usability uncertainties that developers must assess before production deployment.
The free preview lowers the barrier for developers to experiment with a large‑scale, long‑context LLM, a capability that is currently scarce among openly accessible models.
If the promised Apache 2.0 release materializes, the model would become one of the first Chinese‑origin LLMs with a truly permissive, worldwide license, potentially expanding the pool of open‑source models and influencing the global AI research and deployment landscape.
The model’s MoE architecture offers compute efficiency comparable to a 14 billion‑parameter model while retaining the memory footprint of a 280 billion‑parameter model, making it attractive for cost‑sensitive, high‑throughput services.
The lack of EU regional routing and the current restrictive Base license mean that European developers must conduct GDPR compliance assessments before using the API in production, limiting immediate adoption in that market.
Baidu’s emphasis on agent‑style, tool‑integrated tasks signals a strategic push into the emerging market for AI‑driven software development assistants and creative pipelines, potentially reshaping how developers build and test code, 3D assets, and multimedia content.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Key developments to monitor include Baidu’s actual release of the Apache 2.0‑licensed weights, any changes to the pricing model for paid Qianfan usage, and how the model performs in real‑world agent tasks compared to the benchmark claims. Developers should also watch for regulatory scrutiny around GDPR compliance given the lack of EU routing, and for competitive responses from other Chinese labs that are also moving toward open‑weight releases.
Release of the Apache 2.0‑licensed weights by the 25 November deadline and verification of the license file on Hugging Face.
Updates to Baidu’s Qianfan pricing for paid usage beyond the free preview, which could affect the model’s cost competitiveness.
Performance feedback from early adopters, especially on long‑context agent tasks, to see whether real‑world results match Baidu’s benchmark claims.
Regulatory responses or compliance tools that address the lack of EU regional routing for the OpenRouter endpoint.
Competitive moves from other Chinese AI labs (e.g., DeepSeek, Tencent) that are also pursuing open‑weight releases, which could influence the broader open‑source LLM ecosystem.