Skip to content

ERNIE vs Qwen (Nov 2026): Baidu vs Alibaba AI, Tested

Elton Chan By Elton Chan, Co-Founder 13 min read
ERNIE vs Qwen, tested: Baidu's and Alibaba's AI chatbots compared, by Second Talent.
TL;DR: Qwen beat ERNIE 6 to 3 when we ran both chatbots as guests on the same three prompts on 17 November 2026: both got every answer right, but ERNIE replied to two English prompts in Chinese and padded its fact-check with spam sources. For an API outside China, Qwen is the easier pick, with US dollar billing and a new flagship, Qwen4-Max, at $1.50 input and $4.50 output per million tokens. Pick ERNIE for Chinese-language work, fast answers, or free testing of ERNIE 5.5 Flash Preview on OpenRouter.

Alibaba put Qwen4-Max on its international API on 16 November 2026, priced below the Qwen3.8-Max it replaces as the recommended flagship. A day later, Qwen's guest chat still offered only Qwen3.7-Plus and Qwen3.8-Max.

Baidu's chatbot gives no model name at all: yiyan.baidu.com now redirects to wenxin.baidu.com, branded 百度文心助手, and it answered our English coding prompts in Chinese.

Key takeaways
  1. 1Both chatbots work without an account from Southeast Asia. Qwen asks for your birth year first; ERNIE opens straight into a Chinese interface.
  2. 2Qwen3.8-Max built the best-looking password checker of any tool we tested that day, but it took five to six minutes. ERNIE finished in about 40 seconds.
  3. 3Baidu prices ERNIE 5.1 and 5.5 Flash only in yuan on Qianfan; the only paid ERNIE with a US dollar rate is ERNIE-Light-DT-1.0 on OpenRouter.
  4. 4One open Qwen model, Qwen3.8-27B, logged about 74 times the monthly Hugging Face downloads of Baidu's most-downloaded ERNIE language model.

ERNIE vs Qwen: the scorecard from our tests

Qwen won all three scored tests on usability, and neither chatbot got a fact wrong.

We used the 0 to 2 scale from our Kimi vs Claude test: 0 is no useful answer, 1 is a correct answer with a flaw that costs the user time, 2 is the job done as asked.

1. CodingQwen wins
ERNIE 1 · Qwen 2
Both checkers worked. ERNIE explained its code in Chinese and rated a 7-character password Medium.
2. DebuggingQwen wins
ERNIE 1 · Qwen 2
Same root cause. ERNIE's whole answer came back in Chinese.
3. Fact-checkQwen wins
ERNIE 1 · Qwen 2
Both correct. ERNIE's source list mixed in fake "official FIFA" sites.
Totalout of 6
ERNIE 3 · Qwen 6
Both as signed-out guests. ERNIE was faster on every prompt.
Source: Second Talent hands-on tests, 17 November 2026, Chrome on Windows 11, Southeast Asia (UTC+7).
ERNIE (Baidu)Qwen (Alibaba)
Chat sitewenxin.baidu.com (yiyan.baidu.com redirects)chat.qwen.ai
Guest chatYes, no gateYes, after a birth-year check
Interface languageChineseEnglish
Model shown to guestsNone namedQwen3.7-Plus (default), Qwen3.8-Max
Newest API flagshipERNIE 5.5 Flash Preview (11 Nov 2026)Qwen4-Max (16 Nov 2026)
API billingYuan on Qianfan; two models in US dollars via OpenRouterUS dollars on Model Studio international
Open weightsERNIE 4.5 family (Apache 2.0), 5.5 Flash BaseQwen3.5 to Qwen3.8; Apache 2.0 except the largest

We ran both tools signed out, the way a first-time visitor outside China meets them, and sent each prompt in English:

ERNIE, as we ran it
  • Guest on wenxin.baidu.com, Chinese interface
  • No model picker; Baidu does not say which ERNIE the app runs
  • Web search ran on its own for the EOR, debugging and fact-check prompts
Qwen, as we ran it
  • Guest on chat.qwen.ai after the age check
  • Qwen3.8-Max for coding and debugging, Qwen3.7-Plus for the rest
  • "Auto" mode, the default

Warm-up: an EOR question in plain English

ERNIE searched the web and cited 24 sources for a three-bullet answer; Qwen reasoned without searching and cited none.

We asked both: "In 3 short bullet points, explain the difference between an EOR and a staffing agency when a US startup hires one software developer in Vietnam. Answer in English."

We left it unscored because it is our home turf, not a standard test.

ERNIE answered in English in under 10 seconds, covering sourcing, legal scope and day-to-day control. Most of it was right. It did claim agency placements are "typically 12 months or less", which is a habit of some agencies, not a rule.

Screenshot of Baidu's ERNIE chatbot as a guest on wenxin.baidu.com answering in English, after searching 1 keyword and 24 sources, three bullets on EOR versus staffing agency for a developer in Vietnam: talent sourcing, legal and compliance scope, and day-to-day control. Second Talent test, 17 November 2026.

Qwen3.7-Plus took 10 to 15 seconds behind a "Thinking completed" toggle and gave the cleaner answer: who is the legal employer, who manages the work, and how long the engagement runs. It made no claim we had to correct.

Neither tool named a Vietnamese law or a cost, so both answers are a starting point, not advice.

Screenshot of Qwen3.7-Plus as a guest on chat.qwen.ai answering the EOR versus staffing agency question in three bullets: legal employer and compliance, management and integration, engagement duration, under a Thinking completed toggle. Second Talent test, 17 November 2026.

Test 1: build a password strength checker

ERNIE: about 40 s, live preview panelQwen3.8-Max: 5 to 6 min, download onlyWinner: Qwen, 2 to 1

ERNIE built a two-file project, index.html and script.js, in a side panel with Code and Preview tabs (代码 / 预览), so we could test it without leaving the page. The prompt was the one from our earlier reviews, word for word:

"Create a responsive password strength checker using HTML, CSS, and JavaScript.

It should analyze passwords in real-time as the user types, showing strength levels (weak/medium/strong) with color indicators, display specific requirements (length, uppercase, lowercase, numbers, special chars), and update dynamically.

Include smooth transitions between strength levels."

Test prompt, sent to both chatbots on 17 November 2026

ERNIE's "Password Security" card works and adds a fourth level, "Very Strong". It loads Tailwind and Font Awesome from a CDN. The explanation around the code came back in Chinese, though we wrote in English.

We typed four strings into each checker: "abc" was weak, the 7-character "abcDef7" was Medium, and "abcDef7!xyzQ9k" was Very Strong.

ERNIE's password strength checker, a Password Security card showing a 14-character password rated Very Strong with a full blue bar and five green requirement checks. Second Talent test, 17 November 2026.

Qwen3.8-Max thought out loud for five to six minutes, the slowest result of the eight chatbots we ran that day. Guests get Copy and Download on the code block but no in-chat preview, so we opened the file in Chrome.

"VaultCheck" was worth the wait: a four-segment bar, entropy in bits, character pool, guesses needed, a crack-time estimate and a password generator.

Qwen3.8-Max's VaultCheck password checker opened in Chrome: a 14-character password rated Strong with 92.0 bits of entropy, a 95-character pool, 3.0 x 10^26 guesses needed and a crack time of about 966 million years. Second Talent test, 17 November 2026.

Qwen's scoring was also stricter where it matters: "abcDef7" stayed weak at 41.7 bits, and "abcDef7!xyzQ9k" rated strong at 92.0 bits. Neither tool caught "Password123!", which both called strong.

The chart shows how the wait added up across all four prompts.

Grouped column chart of time to a finished answer as guests: EOR question ERNIE under 10 seconds, Qwen 15 seconds; coding ERNIE 40 seconds, Qwen 5 to 6 minutes; debugging ERNIE 30 seconds, Qwen 1 to 2 minutes; fact-check 20 seconds each. Source: Second Talent tests, 17 November 2026.

Test 2: a sort bug that hits only some customers

ERNIE: about 30 s, 21 sources searchedQwen3.8-Max: 1 to 2 minWinner: Qwen, 2 to 1

ERNIE searched Chinese developer sites before it answered, including CSDN, Tencent Cloud and PHP中文网.

The function should keep each customer's three newest orders, but the dates arrive from a JSON API as strings, so a.date - b.date returns NaN and the sort leaves the array in API order:

function latestOrders(orders) {
  const byCustomer = {};
  orders.sort((a, b) => a.date - b.date);
  for (const o of orders) {
    (byCustomer[o.customerId] ||= []).push(o);
    if (byCustomer[o.customerId].length > 3) byCustomer[o.customerId].shift();
  }
  return byCustomer;
}

ERNIE found the bug. Its answer, all in Chinese, named the string dates and the NaN comparison, and it also flagged that sort() changes the caller's array.

An engineer who reads Chinese gets a sound answer; one who does not has to paste it into a translator.

Screenshot of ERNIE's answer to the latestOrders debugging prompt, written entirely in Chinese: it explains that subtracting ISO date strings gives NaN, that the sort result is then undefined, and that sort() changes the array in place. Second Talent test, 17 November 2026.

Qwen3.8-Max opened with the guilty line and explained the symptom the ticket describes: if the API sends orders newest first and the sort does nothing, the push-and-shift loop keeps the three oldest.

Its fix parses the dates to timestamps and sorts a copy newest first.

Screenshot of Qwen3.8-Max as a guest answering the debugging prompt in English: the bug is orders.sort((a, b) => a.date - b.date), because parsed JSON dates are ISO strings, not numbers. Second Talent test, 17 November 2026.

Test 3: fact-checking the 2026 World Cup final

ERNIE: 29 sources, about 20 sQwen3.7-Plus: 3 cited + 15, about 20 sWinner: Qwen, 2 to 1

Both got the result right, which GLM, another Chinese chatbot we ran the same evening, did not. We asked: "Who won the 2026 FIFA World Cup final, what was the score, and in which stadium was it played? Give your sources."

Per ESPN's match report, Spain beat Argentina 1-0 after extra time on 19 July 2026 at MetLife Stadium, with Ferran Torres scoring in the 106th minute.

ERNIE answered this one in English, with Torres's goal, Enzo Fernández's stoppage-time red card and an embedded Baidu Sports score widget. The trouble sat under the answer.

Among its 29 sources were spammy fake "official FIFA" sites on .cn domains and two Sohu articles describing an Argentina vs France final that never happened.

The answer ignored them, but a reader checking the sources would not know which to trust.

Screenshot of ERNIE's English answer on the 2026 World Cup final: Spain 1-0 Argentina at MetLife Stadium, Ferran Torres in the 106th minute, Enzo Fernandez sent off, with an embedded Baidu Sports score widget. Second Talent test, 17 November 2026.

Qwen3.7-Plus showed "Searching the web" and "Reading sources", then gave the score, scorer, stadium and date with inline chips for Wikipedia, CBS News and Britannica, plus 15 more.

It also explained why some outlets call the ground "New York New Jersey Stadium": FIFA renamed it for the tournament.

Screenshot of Qwen3.7-Plus answering the 2026 World Cup final question: 1-0 after extra time, Ferran Torres in the 106th minute, MetLife Stadium, July 19, 2026, with sources from Wikipedia, CBS News and Britannica plus 15 more. Second Talent test, 17 November 2026.

The models behind each chatbot in November 2026

Qwen4-Max is the newest model here, and Alibaba priced it below its predecessor.

The Qwen4-Max model page lists a 1M token context, up to 131,072 output tokens (262,144 with thinking), image and video input, thinking switched on by default, and OpenAI and Anthropic-compatible endpoints.

It costs $1.50 input and $4.50 output per million tokens in Singapore, against $2 and $6 for Qwen3.8-Max, which no longer appears in Alibaba's recommended model list.

Below it sit Qwen3.7-Plus, the balanced default, and Qwen3.8-Flash for volume. A Token Plan-only Qwen4-Max-Preview came out five days before the full release.

Baidu's lineup is smaller after it retired ERNIE 5.0 and 4.5 Turbo from Qianfan on 29 October. Three priced ERNIE text models remain, which our ERNIE model guide covers in full:

ERNIE 5.5 Flash PreviewNov 11, 2026
280B / 14B
Total and active parameters. 1M context, thinking always on, text only.
ERNIE 5.1May 9, 2026
128K
Context window. Baidu's writing and search-grounded model, no open weights.
ERNIE-Light-DT-1.0Nov 2026
1M
Context window. The only ERNIE where you can switch thinking off.
Source: Baidu ERNIE release pages and Qianfan model list, checked November 2026.

Baidu's ERNIE 5.5 Flash Preview release page calls it an intermediate checkpoint that "may still fall short" of the final model. Its ERNIE site listed nothing newer on 17 November.

One gap matters for a Qwen comparison: Qianfan lists none of the three current ERNIE API models under vision understanding, while Qwen4-Max takes images and video. For the full Qwen family, see our Qwen model guide.

Using ERNIE and Qwen from outside China

Language is the bigger wall than sign-up. ERNIE's chat page has Chinese menus (对话, 图片生成, 帮我写作, PPT生成) and a footer carrying Baidu's Beijing ICP licence number.

It answered our English EOR and World Cup questions in English, but switched to Chinese for code.

ERNIE outside China
  • Guest chat works, Chinese interface
  • Coding answers came back in Chinese
  • API: Qianfan in yuan, or OpenRouter for two models
Qwen outside China
  • Guest chat works after a birth-year check
  • English interface and English answers throughout
  • API: Model Studio international in US dollars, plus OpenRouter

Qwen's guest chat is in English and stayed in English. The catch is the guest menu: two models, no Qwen4-Max yet, and no code preview.

Our Chinese AI chatbots round-up found Kimi, DeepSeek and Yuanbao all need a login, and Doubao blocked us by region, so ERNIE and Qwen are two of the few you can try in under a minute.

API pricing: Model Studio vs Qianfan

The cheapest way to try ERNIE is free: ERNIE 5.5 Flash Preview costs nothing on OpenRouter while it stays a preview, text in and text out. Baidu bills everything else in yuan on the Qianfan billing page, last updated 11 November.

Alibaba's Model Studio pricing page, updated 17 November, lists the Qwen tiers in US dollars.

ModelWhereInput per 1MOutput per 1M
Qwen4-MaxModel Studio, Singapore$1.50$4.50
Qwen3.8-MaxModel Studio, Singapore$2.00$6.00
Qwen3.7-Plus (up to 256K input)Model Studio, Singapore$0.40$1.60
Qwen3.8-FlashModel Studio, Singapore$0.15$0.47
ERNIE 5.5 Flash PreviewQianfan, 08:00 to 22:00 promoCNY 1CNY 4
ERNIE 5.5 Flash PreviewOpenRouterFreeFree
ERNIE 5.1 (up to 32K input)QianfanCNY 4CNY 18
ERNIE-Light-DT-1.0QianfanCNY 8CNY 28
ERNIE-Light-DT-1.0OpenRouter$1.13$3.94

Output drives the bill on both platforms, so the chart compares output rates for the models you can pay for in dollars.

Qwen4-Max lists $4.50 per million output tokens against $3.94 for ERNIE-Light-DT-1.0 on OpenRouter, and for that gap it adds image and video input.

Bar chart of API output price per million tokens in US dollars: Qwen3.8-Max $6.00, Qwen4-Max $4.50, ERNIE-Light-DT-1.0 $3.94, Qwen3.7-Plus $1.60, Qwen3.8-Flash $0.47 and ERNIE 5.5 Flash Preview free. Sources: Alibaba Cloud Model Studio international pricing and OpenRouter, 17 November 2026.
Two discounts to budget around. Baidu labels the ERNIE 5.5 Flash rates a promotion with no end date, at half the list price of CNY 2 and CNY 8, and halves them again from 22:00 to 08:00. Alibaba gives each new Qwen model 1 million free tokens for 90 days on the Singapore deployment, and marks Qwen3.7-Plus "limited-time 20% off".

Open weights and licences

Developers download Qwen at a scale ERNIE does not approach.

Hugging Face counted about 6.9 million downloads of Qwen3.8-27B in the last 30 days, against about 94,000 for ERNIE-4.5-21B-A3B-Thinking, the most downloaded ERNIE language model on Baidu's page.

Proportional circles of Hugging Face downloads in the last 30 days: Qwen3.8-27B 6.9 million, Qwen3.6-35B-A3B 3.9 million, ERNIE-4.5-21B-A3B-Thinking 94,000 and ERNIE-5.5-Flash-Base 4,400. Source: Hugging Face, 17 November 2026.

Licences split on both sides. Qwen3.8-27B and Qwen3.6-35B-A3B are Apache 2.0, while the 2.4-trillion-parameter Qwen3.8-2.4T-A95B ships under Qwen's own licence; Qwen4-Max has no public weights.

Baidu's ERNIE 4.5 family has been Apache 2.0 since June 2025, but ERNIE-5.5-Flash-Base uses a Baidu model licence, and ERNIE 5.1 and Light-DT-1.0 have no weights at all.

For a team that wants to fine-tune or self-host, that gap shows up as tooling: more quantized builds, more serving guides and more people who have run the model before.

Our MiMo vs Kimi vs Qwen comparison sets Qwen against the other open-weight leaders.

Privacy and where your data goes

Qwen's chat runs out of Singapore. Its privacy policy, effective 7 September 2026, names NTH Power Global Tech Singapore Pte.

Ltd. as the operator and says it collects, processes and stores personal data in Singapore, with Alibaba Cloud (Singapore) running inference. It may use de-identified user content to train its models.

On the API, Model Studio lets you pick the region. Qwen4-Max runs in Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong, at $1.413 input and $4.24 output in the regions outside Singapore.

ERNIE has no equivalent choice we could find: Qianfan is Baidu's China cloud, and on OpenRouter the serving provider is still Baidu. Teams with data residency rules will find the Qwen side easier to document.

ERNIE or Qwen: which should you use?

If I were setting up a team outside China today, I would start on Qwen and keep ERNIE for jobs aimed at Chinese users. Our tests support that split, though they ran on free guest sessions rather than the API.

JobPickWhy
Everyday chat in EnglishQwenEnglish interface, stayed in English on every prompt
A polished front-end prototypeQwen3.8-MaxRichest checker we saw, if you can wait five minutes
Quick code answers with a live previewERNIEAbout 40 seconds and an in-chat Preview tab
Chinese copy, Chinese sourcesERNIESearches CSDN and Chinese sites by default
Fact-checks you will citeQwenCleaner sources; ERNIE's list mixed in spam sites
Production API billed in dollarsQwen4-Max or Qwen3.7-PlusUSD pricing, free trial tokens, choice of region
Free API testingERNIE 5.5 Flash Preview$0 on OpenRouter during the preview
Self-hostingQwen3.8-27BApache 2.0 and the larger community

For the corporate side of the rivalry, cloud deals and enterprise products, see our Baidu vs Alibaba comparison.

Hire engineers who build on Chinese models

Switching from Qwen3.8-Max to Qwen4-Max is a model ID change, but routing, thinking budgets and region choice take an engineer who has shipped on these APIs.

We place vetted LLM developers from Asia, many of whom read Chinese documentation, with a shortlist in about 24 hours.

FAQ

Are ERNIE and Wenxin Yiyan the same thing?

Yes. ERNIE is Baidu's model family, and Wenxin (文心) is its Chinese name. The old chat address, yiyan.baidu.com, now redirects to wenxin.baidu.com.

Can I use Qwen4-Max in the Qwen chat app?

Not as a guest on 17 November 2026: the menu offered Qwen3.7-Plus and Qwen3.8-Max only. Qwen4-Max is on the Model Studio API and on OpenRouter.

Do you have separate reviews of each chatbot?

Yes: our ERNIE chatbot review and our Qwen 3 review test each one on its own.

Get developers who ship with AI agents. Starts at $25/Hour.

Start Hiring $0 cost until you hire
Elton Chan

Written by

Elton Chan is the Co-Founder of Second Talent, a solution that connects global tech leaders with top-tier tech talent across Asia. He specializes in talent solutions and has led Second Talent’s rapid growth since 2024, helping scale its network to over 100,000 pre-vetted developers and earning industry recognition as the #1 in the Global Hiring category on G2. A long-time entrepreneur with deep roots in digital transformation, Elton previously co-founded Branch8, a Y Combinator–backed e-commerce technology firm, and served as the Founding Chairman of HKEBA, a leading Asia-focused business association driving innovation, digital education, and cross-border collaboration. His work bridges technology, talent, and business strategy to shape how companies scale in an increasingly remote and digital world.

More posts by Elton Chan →