ERNIE 5.5 Flash Preview launched in four acts: the base weights on November 5, Preview on November 11, the Baidu World stage on November 12, and an open-source promise for November 25 that has not yet been kept. Each act has its own audience — act one for researchers, act two for users, act three for the market, act four still pending. This article assembles everything confirmed across the four acts, and states what the model is: the first post-trained version of ERNIE 5.5 Flash Base, not a new base model; and Preview is an intermediate checkpoint in post-training, not the final release.
On November 5, 2026, ERNIE 5.5 Flash Base released: the pre-training endpoint weights landed on Hugging Face (baidu/ERNIE-5.5-Flash-Base, FP8 and BF16 builds) under Baidu's custom "ERNIE 5.5 Flash Base Model License Agreement v1.0," not Apache 2.0. Base is base-model capability only, with no instruction following — downloadable, but not directly chattable. Pre-training data: 33.5 trillion tokens; 128K training context, fine-tuning/inference supports 1M. This act matters as an anchor: the Preview released six days later is not a new model appearing from nowhere, but the first post-trained version of this base — keeping November 5 in the timeline is what separates "base release" from "product release" as two different events.
On November 11, 2026, Baidu officially released ERNIE 5.5 Flash Preview: the first post-trained version of Base. The official release page (ernie.baidu.com) is dated Nov 11, the official WeChat post went out the same day, and 每日经济新闻, 新华网, 新智元 and 量子位 all covered it the same day. The specs: a 280B-total / 14B-active sparse MoE, 1M-token context, text in and text out, fixed high-effort reasoning with no toggle, tool calling and structured output; techniques are hybrid sparse attention + multi-router MoE + ECHO multi-token prediction, plus the first use of Baidu's in-house RSI (recursive self-improvement), with ERNIE Evolve contributing to inference optimization. Two lead-in signals the same week: on Nov 9 Baidu first disclosed its RSI progress, and on Nov 10 OpenRouter listed the paid and free tiers early — Preview was already callable via API a day before the official release. The release page also carried the ~20-row vendor benchmark table (all self-reported; see the benchmarks article) and Qianfan pricing.
On November 12, 2026, Baidu World 2026 opened in Beijing, Robin Li keynoted, and ERNIE 5.5 Flash Preview shared the stage with the Token Exchange (词元所), a marketplace for model services. This was the launch's second wave of exposure: wave one was the Nov 11 release page and press coverage, wave two the main conference stage. The vendor says the Token Exchange signed 100+ vendors in its first month, with 50M calls and 700B tokens per day — company-reported figures with no independent audit; this site reports them, it does not vouch for them. The stage also carried one key statement: Sun Tianxiang, head of Baidu's foundation model R&D, said the team is working to "drop the Preview suffix as soon as possible" — the final version follows, but no timeline was given.
Act four has not happened. The vendor promises the Preview weights will be open-sourced around November 25, but the wording splits in two: the English release page says "by Nov 25" (read directly by digintoai), the Chinese WeChat post says "formally released to the global open-source community on Nov 25" (via aiho.net). Whether it is "that day" or "by that day" can only be settled once the weights actually land. As of November 16, Hugging Face hosts only the Base repo — no Preview repo; and the license accounts conflict (see the deployment article). In a timeline, the distinction matters: November 25 is a promise, not a fact, and writing it as "already open-sourced" is wrong.
Preview is not a new base model. It is the first post-trained version of ERNIE 5.5 Flash Base (the Nov 5, 2026 pre-training endpoint weights): the same 280B-total / 14B-active sparse MoE, the same 33.5 trillion tokens of pre-training data, with post-training turning it from "base capability only" into a product that can chat, be called via API and run agents. Its positioning is the execution layer for agent workloads: coding, 3D modeling, knowledge work and content creation. But Baidu's own words draw the boundary at the same time: Preview is an intermediate checkpoint in post-training that "may have capability gaps due to insufficient training" — which means every benchmark score on the Nov 11 release page is intermediate, and the final version may change them. Read every score on this topic with that premise attached.
The official demos deserve a paragraph of their own, because they are easy to mistake for benchmarks. The third-person wuxia game 《云岫行》 (UE 5.8 + C++, completed in a single 11-hour agent session), ~100 phone mini-games generated from design documents, and airport/airliner Blender modeling (dual-agent) are all end-to-end results inside Baidu's own agent framework — a different harness or prompting strategy could change results by an order of magnitude; and the model is text-only, so it cannot see the screen — visual problems in the 3D and game demos were solved by transcribing them into code tests. madechango / 火星财经's Nov 13 hands-on verdict, "good taste, strong coding," is third-party hands-on reporting and should be read separately from the official demos.
Day-one reaction split two ways: Chinese tier-1 tech media were hot — 新智元 published a launch-day deep dive on Nov 11, 量子位, 每日经济新闻 and 新华网 followed the same day, and IT之家, 雷锋网 and QbitAI covered Baidu World; English was cold — Unite.AI, aiaffairs, artificialwatch, singularity.kiwi, freshsynapse, aireiter, apidog and digintoai covered it, but as of Nov 16 no TechCrunch, The Verge, Reuters, Bloomberg or Wired piece has surfaced. That split is itself the ERNIE 5.5 Flash Preview profile: the strongest Chinese coverage of this discovery round, with English reach still the weak spot; search interest matches — English has bare-term autocomplete with zero long-tail suffixes, and Chinese has zero across the board.
November 11, 2026 (official release page and WeChat post the same day); featured at Baidu World 2026 on November 12. Note the base model, ERNIE 5.5 Flash Base, came earlier — November 5 on Hugging Face; November 25 is a pending open-source promise, not a release date.
Preview is an intermediate checkpoint in post-training, self-described as unfinished: it "may have capability gaps due to insufficient training." The final version follows, and Sun Tianxiang, head of Baidu's foundation model R&D, said the team is working to "drop the Preview suffix as soon as possible" — but no timeline was given.
Yes. The Qianfan API and OpenRouter (paid and free tiers) were live, the web chat at ernie.baidu.com/web/new-chat was open, and OpenRouter had even listed it early on November 10. What you could not use was the weights — Preview weights are not public as of November 16, with open-sourcing promised for around November 25.
Reference: Baidu official release page — ERNIE 5.5 Flash Preview · PR Newswire release · OpenRouter — baidu/ernie-5.5-flash-preview:free (rewritten and organized by ModelBrief; facts per the official release page and reporting; the official page could not be read directly this round, facts cross-checked via secondary sources)