The Base–Preview relationship is one sentence: the same 280B base, with Preview as its first post-trained version. But behind that sentence are three differences — capability, weights and license. Base can be downloaded but not chatted with; Preview can be chatted with and called via API, but its weights are not out and its license is unsettled. This article states all three differences at once.
ERNIE 5.5 Flash Base is the pre-training endpoint weight released November 5, 2026: a 280B-total / 14B-active sparse MoE, pre-trained on 33.5 trillion tokens, with a 128K training context and fine-tuning/inference support for 1M. ERNIE 5.5 Flash Preview is the first post-trained version of Base, released November 11, 2026: the same base, the same parameter count, with post-training turning it from "base capability only" into a product that can chat, be called via API and run agents. The positioning changes with it: Base is weights in a researcher's hands; Preview is the execution layer for agent workloads (coding, 3D modeling, knowledge work, content creation).
| Dimension | ERNIE 5.5 Flash Preview | ERNIE 5.5 Flash Base |
|---|---|---|
| Nature | Base's first post-trained version; an intermediate post-training checkpoint | Pre-training endpoint weights (the base model) |
| Released | Nov 11, 2026 | Nov 5, 2026 |
| Instruction following | Yes (a post-trained product) | No (base capability only) |
| Reasoning tier | Fixed high-effort tier, no toggle | — |
| Tool calling / structured output | Supported | — (base capability; post-train it yourself) |
| Modality | Text only (text in / text out) | Text only (base model) |
| Benchmarks | ~20-row vendor table on the official release page (all self-reported) | No official benchmark table |
| Techniques | Hybrid sparse attention + multi-router MoE + ECHO multi-token prediction; RSI, ERNIE Evolve | Same base techniques |
One detail worth noting: Preview's reasoning is a fixed high-effort tier with no toggle (aireiter's analysis) — a behavioral trait only a post-trained version has; as a base model, Base has no "tier" concept at all. Likewise, the official demos (the 《云岫行》 game built in one 11-hour agent session, ~100 phone mini-games) are Preview's post-trained capabilities, not Base's; Base has no official benchmark table, and any claim about Base scores has to answer "who ran it, on what harness" first.
Base weights are public: Hugging Face repo baidu/ERNIE-5.5-Flash-Base, FP8 and BF16 builds, downloadable. Preview weights are not public as of November 16 — there is no Preview repo on Hugging Face; the vendor promises open-sourcing around November 25, but the English page says "by Nov 25" and the Chinese post says "on Nov 25," and the wording split can only be settled once the weights actually land. In short: today, only the base can be downloaded; to use Preview, go through the API (Qianfan or OpenRouter) or the web chat.
Base's license is settled: Baidu's custom "ERNIE 5.5 Flash Base Model License Agreement v1.0," not Apache 2.0, with restricted commercial terms — freshsynapse says the license excludes the EU, UK and Korea, apidog says non-commercial use only, and the details differ, so the unified wording is "a custom license with restricted commercial terms; see the license file on Hugging Face for details." Preview's license is unsettled: several secondary sources (digintoai, singularity.kiwi, yunsilu, saeedai) say Apache 2.0, but aiho.net — which read the official release page directly — states plainly that "the license type has not yet been stated by the vendor." The conclusion: Apache 2.0 is not confirmed by the official release page and must not be stated as fact — write "several outlets say Apache 2.0; the official release page does not confirm it; verify against the license file when the weights land on Nov 25." One reminder: Base's custom license is not Preview's license; wait for the Nov 25 license file itself — do not infer Preview's license from Base's, and do not infer Base's from the rumored Apache 2.0.
To do your own post-training or research, or to have weights in hand: download Base (FP8 + BF16 public), but read the Hugging Face license file first — commercial terms are restricted. To chat, call an API or run agent tasks directly: use Preview (Qianfan API, OpenRouter paid and free tiers, web chat all live), accepting the Preview intermediate state and the weights not being out. The two are not substitutes: Base is the base model, Preview is the base model's first post-trained version — the choice depends on whether you want "weights" or "a product."
Same base, two products. Base is the Nov 5 pre-training endpoint weight (base capability only, no instruction following); Preview is Base's first post-trained version (Nov 11), able to chat and be called via API. Same parameter count (280B/14B), different capability and availability.
No. Base has base capability only, with no instruction following; chatting with it needs your own post-training or prompt engineering, with unvalidated results. To chat directly, use Preview (API or web chat).
Unknown. Base carries a custom Baidu license (not Apache 2.0); Preview's license is not stated on the official release page, several outlets say Apache 2.0, and the license file must be checked when the weights land on Nov 25. Do not infer Preview's license from Base's.
Reference: Baidu official release page — ERNIE 5.5 Flash Preview · Hugging Face — baidu/ERNIE-5.5-Flash-Base · aiho.net — 280B deep dive (rewritten and organized by ModelBrief; facts per the official release page and the Hugging Face model page)