Baidu has released a new artificial intelligence model that is big, fast, cheap — and, if the company keeps its promise, about to become open source. Baidu ERNIE 5.5 Flash Preview, released on November 11, 2026 and presented at the Baidu World 2026 conference in Beijing on November 12, is a 280-billion-parameter model with a one-million-token context window. Baidu says it will publish the model’s weights before November 25, 2026, under the Apache 2.0 licence.
That date is what makes this launch worth watching: a frontier-scale model with open weights would let developers download, study and run a powerful system themselves instead of renting API access. It is also the latest move in the busiest week of AI launches this year, arriving within days of Google’s Gemini 4 Hydrogen and Anthropic’s Claude Fable 5.2. The race is no longer only about the smartest model, but the fastest, cheapest and most open one.
What Is Baidu ERNIE 5.5 Flash Preview?
Baidu ERNIE 5.5 Flash Preview is the first post-trained version of ERNIE 5.5 Flash Base, a raw pre-training checkpoint Baidu published on November 5, 2026. The Preview is the version trained to act as a chat and agent model, and Baidu describes it as an intermediate checkpoint, with post-training still under way and an official version to follow.
The model is a sparse mixture-of-experts design: it holds a huge number of parameters but only activates a small slice of them for each token it generates. It is a text-only reasoning model, so it reads and writes text rather than seeing images or video, and it works through problems step by step, with that reasoning fixed at a high effort level that cannot be switched off.
Baidu also announced two application models — ERNIE Wizshot Micro-Expression, which creates video from text, image and speech, and ERNIE Light Deep Thinking, built for complex requests and tool calls — plus a Token Exchange marketplace for buying model services. Baidu says its ERNIE Evolve technology makes running the model more efficient.
The Headline Numbers: 280B Parameters and a 1M-Token Context
The figures Baidu published are striking. The model has 280 billion total parameters, with only 14 billion active per token, and a context window of one million tokens, enough room to hold a very large codebase or a stack of long documents in a single session.
Speed is the headline claim. Baidu says the model generates more than 200 tokens per second on its Qianfan platform. Independent measurements so far tell a more modest story: OpenRouter, which hosts the model, reports a median of about 90 tokens per second with 2.38 seconds of median latency.
The price is low by frontier-model standards. On Baidu’s Qianfan platform the model costs 1 yuan per million input tokens, roughly 15 US cents, and 4 yuan per million output tokens, roughly 60 cents. Those rates are currently at a limited-time 50 percent discount, and they double during busy periods. For developers who only need to test the model, the price is zero: OpenRouter serves it free, with limits of 20 requests per minute and 50 requests per day, rising to 1,000 per day once an account has bought 10 dollars of credits.

What Baidu Says It Can Do: Coding, 3D Modelling and Long Documents
Baidu is pitching the model first and foremost at developers. The company says it prepared 14 million executable code tasks for training, and it reports a score of 1,668 on CodeArena WebDev, a test of building websites from prompts, alongside competitive results on Terminal-Bench and AutomationBench. All of those results are Baidu’s own claims, and no independent benchmark of the Preview has been published yet.
The demonstrations Baidu showed are ambitious. In one, the model spent 11 hours building a wuxia game in Unreal Engine 5.8 and produced around 100 mobile mini-games. In 3D work it built an airport and an airliner with the Blender Python API and reconstructed the Horyu-ji temple complex with about 2,100 objects and 1.82 million triangles. For knowledge work, Baidu showed 20- and 24-page research reports and an original orchestral piece lasting 6 minutes and 34 seconds.
Those demos deserve a careful reading: Baidu’s own materials credit external tools and third-party assets in the finished work, so they show what the model can do inside a well-built agent system, not what it produces on its own.
How to Try Baidu ERNIE 5.5 Flash Preview Free Today
You do not need to wait for the open weights to try the model. The API is already live in two places.
The first route is Baidu’s own Qianfan platform, where the model ID is ernie-5.5-flash-preview. You need a Baidu AI Cloud account, and calls are billed to your balance in yuan. The second route is OpenRouter, where the model is listed as baidu/ernie-5.5-flash-preview:free and costs nothing per token within the free-tier limits. For most readers in the United States, the United Kingdom and Canada, that is the easiest way to test it: create an OpenRouter account, generate an API key, and send a first request with the free model ID, including the :free suffix.
There are practical limits to know before you build on it. Because the model is a text-only reasoning model, it cannot see images or screens, and its reasoning step cannot be turned off or dialled down, so every call pays reasoning tokens. The free tier is also rate-limited, and free tiers change quickly, so check the provider’s own page before relying on one.
The Open-Source Promise — and the Licence Catch
The most important date in this story is November 25, 2026. That is when Baidu says the Preview weights will be open-sourced, and its English product page pledges the Apache 2.0 licence. If that happens, developers will be able to download the model, run it on their own hardware and build products on top of it without paying Baidu per token.
But the promise is not yet a fact. Baidu’s press release names no licence and sets only a deadline of November 25, while the product page pledges Apache 2.0 and says only “soon”, without a date. The only comparable release so far, the ERNIE 5.5 Flash Base model published on November 5, carries a custom licence that excludes the European Union, the United Kingdom and South Korea, and requires written permission for commercial use above 50 million dollars in revenue. For readers in the UK, that exclusion is directly relevant: the Base weights cannot legally be used there, and only a genuine Apache 2.0 release of the Preview would change that.
There are practical hurdles too. The weights are not on Hugging Face yet — only the Base checkpoint and its BF16 variant are listed — and no repository for the Preview exists. Independent estimates put the FP8 weights at about 288 gigabytes, and the llama.cpp project has not yet merged support for the architecture, so running it locally on a laptop is not an option today. The honest summary is that Baidu ERNIE 5.5 Flash Preview is open-weight in promise, not yet in practice, and November 25 is the date that will settle it.

How Baidu ERNIE 5.5 Flash Preview Stacks Up Against Gemini 4 Hydrogen and Claude Fable 5.2
This launch lands in a remarkable week for AI. Anthropic released Claude Fable 5.2 on November 9, a flagship model that coordinated more than 1,000 AI agents over eight days to build a working operating system, now costs 8 dollars per million input tokens and 40 dollars per million output tokens. Google followed with Gemini 4 Hydrogen on November 12, a fast execution model that completes tasks 2.5 times faster than Gemini 3.8 Flash and costs 1.50 dollars per million input tokens and 7.50 dollars per million output tokens.
Against that field, Baidu’s model competes on price and openness rather than raw capability. Its Qianfan pricing undercuts Hydrogen by roughly ten times on input tokens, and its planned open weights would make it the most accessible frontier-scale model of the group. What it cannot yet claim is verified performance: Fable 5.2 leads the Artificial Analysis Intelligence Index at about 65, and Hydrogen scores 51, while ERNIE 5.5 Flash Preview has no independent index score at all, only Baidu’s own benchmark rows.
What Happens Next
Three things will decide whether this launch matters or fades: whether the weights appear on Hugging Face on November 25 with a licence file that really says Apache 2.0; whether independent benchmarks confirm Baidu’s coding and speed claims; and whether the free tier on OpenRouter turns into real projects.
For now, the sensible move is to try the model free and judge it on your own work: send it a coding task you already know the answer to, and compare the result with the models you already use. Launch-week claims are cheap; a model earns its place by doing your job well, repeatedly, at a price you can afford. Baidu ERNIE 5.5 Flash Preview has made a strong opening case, and the next two weeks will show whether the substance matches the pitch.
