Baidu's ERNIE 5.5 Flash Preview Is Live on Qianfan — Weights Open 25 November

Baidu unveiled ERNIE 5.5 Flash Preview at Baidu World 2026 on 12 November — and this one you can actually use today. The API is live on Baidu AI Cloud's Qianfan platform, and the model is also available for free on OpenRouter. The weights follow on 25 November.

Before the details: this is not the model we covered on 5 November. That was ERNIE 5.5 Flash Base — a raw, pre-training-only checkpoint for researchers and fine-tuners. This is the Preview of the post-trained model: the one that follows instructions, writes code, and does the work. Same 280-billion-parameter MoE underneath (14 billion activated per token), same 1-million-token context ceiling — but a different thing to use, and a different set of questions to ask about it.

What you can do with it today

The API is live on Qianfan right now, priced at RMB 1 per million input tokens and RMB 4 per million output tokens, with cached input at just RMB 0.04 per million tokens. That cached-input price is the number that matters for agent workloads — long tool-call loops and multi-turn conversations live or die on cache pricing, and RMB 0.04 is aggressive.

It is also free on OpenRouter, which is the practical route for most developers outside China: no Qianfan account, no RMB billing, just an API key. That is where I would try it first.

Baidu says the model is strong in coding, 3D modeling, knowledge work and content creation, and competitive on Terminal-Bench and AutomationBench while decoding very fast. Those two benchmarks are worth noting because they are the ones that actually resemble developer work — Terminal-Bench is real terminal tasks, AutomationBench is multi-step automation — rather than exam-style quizzes. But both numbers are Baidu's own, from Baidu's own release. No independent lab has run them yet, and "competitive" is doing a lot of work in that sentence without the actual scores attached. Treat the benchmark claim as a reason to test the model, not as a result.

Under the hood, Baidu credits ERNIE Evolve, its proprietary recursive self-improvement technology, for the inference efficiency. That is a claim about a training method, not something you can verify from the API. What you can verify is the decoding speed and the price on your own workload — and that is the honest test.

What is actually new here (and what is not)

The Preview is the headline, but Baidu World was a full-stack announcement. Here is the rest, in the order a developer should care about it:

• ERNIE Wizshot Micro-Expression — an application model for highly expressive video creation. It processes text, images and speech together and handles long-form text input of up to 4,000 Chinese characters. Baidu says it leads the industry for podcasts and short animated dramas at half the typical cost. If you make video content, this is the one to watch — but "leads the industry" is Baidu's claim, and video quality is the hardest thing to judge from a press release. Wait for real samples.
• ERNIE Light Deep Thinking — built for complex requests and tool calls at search-scale speed. It pairs a fast-thinking model that serves users with a slow-thinking model that continuously refines it in a closed loop, and it uses Baidu Sparse Attention to compress historical memory during decoding without hurting performance. The fast/slow pairing is a sensible design for search and agent work; whether it holds up under real tool-call loads is untested outside Baidu.
• Token Exchange — a marketplace connecting model service providers to customers, with side-by-side comparison of pricing and performance data. It verifies resources by comparing model fingerprints and testing capability, performance and availability, and rates providers with spot checks and end-to-end call monitoring. One month in, it has onboarded more than 100 model resources and handles a daily average of over 50 million calls and 700 billion tokens. A neutral marketplace for model tokens is a genuinely interesting idea — the model-API market is fragmented and opaque, and a trusted exchange could lower the cost of switching providers. But it is Baidu's exchange, and "neutral" is a claim that needs a year of operation to prove.
• Agent upgrades — DuMate is now the unified gateway to Baidu's specialised agents, with its Harness engine and Agent API fully opened; the team says it iterated more than 160 times in six months, with 90% of its code and 50% of its external communication materials generated by DuMate itself. KOOKO, upgraded from GenFlow 3.0, is a document and creative workspace drawing on 1.8 billion professional resources and 750 million academic papers. Baidu Yijing, the digital human agent, now does short animated drama creation end to end — concept to final cut — at RMB 0.2 per second of video. Famou 3.0, the self-evolving agent for industrial operations, now spans 25 of China's 41 industrial categories.
• Chips and cloud — the Tianchi 1024-card Supernode, an all-optical interconnect supernode; Kunlunxin's M100 inference chip has shipped as planned, and the flagship M300 is expected in 2027. Robin Li's framing for all of this was the "intelligent organization" — Baidu's pitch is that AI's next chapter is not better chatbots but companies rebuilt around tasks, with agents completing more than a million tasks a week inside Baidu itself, up from 84,000 a year ago.

The honest caveats

• Preview means preview. The weights are not out yet — 25 November is the date Baidu gave. Until then you are using a hosted API, and hosted APIs change.
• The licence for the weights is unknown. This matters more than it sounds. The 5 November Base model shipped under a custom licence that excluded the EU, UK and South Korea and gated commercial use above $50 million in related revenue behind Baidu's written permission. Do not assume the Preview weights will carry the same terms — and do not assume they will be Apache 2.0 either. Read the licence on 25 November before you build anything on the weights.
• Qianfan is the home platform. The OpenRouter listing makes it easy to try, but the primary API, the application models and Token Exchange all live on Baidu AI Cloud. Developers outside China should expect friction: account setup, RMB billing, and documentation that is not always English-first.
• All benchmarks are Baidu's own. Terminal-Bench and AutomationBench scores are described as "competitive" with no figures published in the release. The MMLU 93.3 and HumanEval+ 82.3 from the Base model's table were Baidu's own too. Independent runs are the only thing that turns these claims into facts.
• The price is in RMB and the free tier is OpenRouter's, not Baidu's. Pricing on Qianfan is genuinely cheap, especially cached input — but check what the OpenRouter free listing actually rate-limits before you build a demo on it.
• No hands-on testing was performed for this article. Everything here comes from Baidu's press release as published on 12 November 2026.

What I would actually do

Try it on OpenRouter today, on a real task — not a benchmark, a task. Give it a multi-file coding job with a long context: a repository you know well, a refactor you have already done by hand, so you can tell whether the output is good or merely fluent. Run the same job through whatever you currently use and compare the cost, not just the quality. The RMB 0.04 cached-input price only matters if the model is good enough that you keep it in the loop.

If you fine-tuned or studied the Base model, the Preview is the payoff for that work — the post-trained version is the one most developers actually wanted. But wait for 25 November and read the licence before you commit to the weights. The Base licence was the real story of that release, and it would be a mistake to assume this one is different — in either direction — until the text is public.

And if you are watching the open-weight race rather than using the model: the pattern to notice is that the race is no longer about parameter count. It is about inference cost, cache pricing, licence terms, and whether a model is actually usable the day it is announced. On that last measure, ERNIE 5.5 Flash Preview scores well — it is live today, not "coming soon." The rest is still to be verified.

Have you tried it yet? If you run it against your own workload, I would genuinely like to hear whether the decoding speed and the long-context behaviour hold up — that is the part no press release can prove.

Sources: Baidu press release via PRNewswire, "Baidu Unveils Latest ERNIE Models Amid Broader Full-Stack AI Lineup, and Debuts Token Exchange to Power Intelligent Organizations" (12 November 2026, published 22:29+08:00); Manila Times mirror of the same release. Pricing, benchmark characterisations and agent claims are Baidu's own. Our earlier coverage of the 5 November base model: "Baidu's ERNIE 5.5 Flash Base: A 280B Open-Weight Model You Can Fine-Tune — Explained Honestly". No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.