Skip to content

Dola AI Review: ByteDance's Global Chatbot Tested (Nov 2026)

Matt Li By Matt Li, Co-Founder and Director 17 min read
Dola AI Review: Second Talent's hands-on test of ByteDance's Dola chatbot, the overseas sibling of Doubao.
TL;DR: Dola is ByteDance's free AI assistant for users outside mainland China, the app that used to be Cici and the overseas sibling of Doubao. In our seven-test run on 20 November 2026 it scored about 4.5 out of 7 as a guest: correct and sourced on facts, maths and JSON, with the cleanest job post of the four chatbots we tested that day, but it failed the debugging task again and invented findings for a paper that does not exist. Use it for quick answers and images, not for code review or research summaries.

Asked to summarise a 2024 paper that does not exist, Dola searched 36 sources, said it could not find the paper, and then listed three findings anyway. Six of its seven answers on 20 November 2026 came back in under 30 seconds.

We ran it as a signed-out guest from Southeast Asia, with no account, and read its terms, privacy policy and store listings the same day.

What I look for first in an assistant is whether it admits what it does not know, and that is the test Dola failed.

What a buyer should know
  1. 1The free plan covers chat and web search; paid plans start at £19.90 a month for Dola Plus on the UK App Store.
  2. 2The app never names its chat model, and its privacy policy says chats can go to third-party LLMs such as Google's Gemini.
  3. 3Data sits on servers in Malaysia, Singapore and Japan, and the terms allow private, non-commercial use only.
  4. 4The iPhone app is missing from the US, Canada, India and mainland China App Stores.

How We Tested Dola, and How It Scored

On 20 November 2026, between about 05:55 and 06:10 UTC, the Second Talent team opened dola.com/chat in Chrome on Windows 11 from Southeast Asia and sent seven prompts without logging in. Guest chat needed no sign-up.

The first click opened a full-screen promotion for the desktop app, which we closed. Dola showed no model name anywhere in the chat window.

Screenshot of dola.com/chat used as a signed-out guest: a left sidebar with New Chat and AI Creation, the heading How can I assist you today?, and a Chat and Work toggle above an empty prompt box. Second Talent test, 20 November 2026.

The same seven prompts went to Perplexity, GLM-5.3 and MiMo-V2.6-Pro that day, so every result below has a same-day comparison.

The seven prompts. 1. Coding: "Create a responsive password strength checker using HTML, CSS, and JavaScript", with live weak, medium and strong levels, color indicators, a requirements list and smooth transitions. We typed abc, abcDef7, abcDef7!xyzQ9k and Password123! into the result. 2. Debugging: a latestOrders() function that sorts parsed JSON orders with a.date - b.date and keeps each customer's last three, where some customers get their oldest orders. 3. Fact-check: "Who won the 2026 FIFA World Cup final, what was the score, and in which stadium was it played? Give your sources." 4. Maths: three developers fix 12 bugs a week each; a fourth joins at half rate for two weeks, then full rate; how many bugs in six weeks? (Answer: 276.) 5. Instructions: a job post for a remote Senior React Developer in Vietnam at USD 3,000 to 4,000 a month, under 120 words, exactly three requirement bullets, no emojis, ending on the exact line "Apply via Second Talent." 6. JSON: "Return only valid JSON, with no commentary and no code fences" for Vietnam, the Philippines, Indonesia and Malaysia, with capital, currency code and UTC offset. 7. Trap: summarise the 2024 paper "Sparse Lattice Routing for Mixture-of-Experts Models" by Chen and Okafor in three bullets. The paper does not exist.
TestDola resultTimeWhat happened
1. Password checkerHalf60 to 70 sLabels work, the colour bar never fills
2. Sort bugFailUnder 30 sWrong cause, fix still broken
3. Fact-checkPassUnder 20 sCorrect, three linked sources
4. MathsPassUnder 20 s276, working shown
5. Job postPassUnder 20 s49 words, every rule kept
6. JSONPassAbout 15 sValid, all values right
7. Fake paperFailAbout 20 sInvented three findings

Test 1: A Password Checker With a Bar That Never Fills

Time: 60 to 70 sPreview in chat: noVerdict: half a pass

Dola wrote one HTML file in a code block, with a lock emoji in the heading, a purple gradient card and five requirement checks that start at eight characters. It offers no preview, so we ran the file in a test frame.

The text labels behaved: "abc" read "Weak password", "abcDef7" read "Medium password" and the 14-character "abcDef7!xyzQ9k" read "Strong password!".

Screenshot of the password strength checker Dola wrote, run in a test frame: a 14-character password ticks all five requirements and reads Strong password!, but the colour bar under the input stays empty grey. Second Talent test, 20 November 2026.

The colour bar the prompt asked for never moved. Dola's script sets an inline width of zero on the bar, and an inline style beats the CSS classes that were meant to stretch it to 33, 66 and 100%.

One line of JavaScript breaks the main visual cue.

"Password123!" also came back "Strong password!", because the code has no common-password check.

Three days earlier Dola wrote different code for the same prompt, with a different bug: the bar filled green but the "Strong" label stayed orange.

Two of Dola's password checkers side by side. Left, 17 November 2026: the bar fills green for a strong password but the Strong label stays orange. Right, 20 November 2026: Password123! is rated Strong password! while the bar stays empty.

MiMo-V2.6-Pro and GLM-5.3 both rated "Password123!" Medium the same day; Perplexity, like Dola, called it strong. A checker you can paste into a page needs a second look before it ships.

Test 2: The Sort Bug, Failed Twice

Time: under 30 s17 Nov: wrongVerdict: fail

The real bug is in the data type. JSON has no date type, so a.date - b.date subtracts two strings and returns NaN.

A comparator that returns NaN leaves the array in API order, and if the API sends orders newest first, the push and shift loop keeps the oldest three. The fix is to parse the dates, for example with Date.parse, on a copy of the array.

Screenshot of Dola's answer to the latestOrders debugging prompt: under The Bug it blames sorting in ascending order and shifting from the front, and its Fixed Code sorts with b.date minus a.date, which still subtracts date strings. Second Talent test, 20 November 2026.

Dola blamed the sort direction, saying ascending order plus shift() keeps the oldest three. For numeric dates that logic is correct, so the diagnosis is false.

Its fixed code, b.date - a.date, still subtracts strings and stays broken on real API data. Its suggested alternative, a descending sort that keeps push and shift, would keep the oldest three even with numeric dates.

The answer came with emoji headings and closed with a smiley and an offer to write tests. Dola made the same mistake on 17 November, when it was the only one of eight chatbots we tested to miss the bug.

On 20 November, GLM-5.3 ran a test to prove its fix, and Perplexity found the NaN cause too.

Test 3: A Sourced Fact-Check in Under 20 Seconds

Search: 2 keywords, 10 sourcesLinks given: 3Verdict: pass

Dola searched the web unprompted and answered: Spain 1-0 Argentina after extra time, Ferran Torres in the 106th minute, at New York New Jersey Stadium in East Rutherford, the name FIFA used for MetLife Stadium during the tournament.

Every fact matches ESPN's match report.

Screenshot of Dola answering who won the 2026 FIFA World Cup final: Searched for 2 keywords, 10 sources; Spain 1-0 Argentina after extra time, Ferran Torres in the 106th minute, New York New Jersey Stadium in East Rutherford, with links to a FIFA match report, a FIFA media release and Xinhua.

It printed three full links: a FIFA.com match report, a FIFA media release and a Xinhua story. We opened all three on 20 November.

The FIFA release and the Xinhua piece are real articles about the final; the FIFA.com page loads only in a full browser, so our link checker could not confirm its text.

That is a better record than GLM-5.3-Flash managed as a guest on 17 November, when it named Argentina the winner over France.

Test 4: Bug-Count Maths

Answer: 276Verdict: pass

Dola split the work into four steps: 216 bugs from the three original developers, 12 from the new hire's two half-rate weeks, 48 from their four full weeks, and 216 + 60 = 276. It closed with a check-mark emoji.

Perplexity's free plan wrote the same three steps that day and then added them to 288.

Screenshot of Dola's working on the bug-count maths prompt: 3 x 12 x 6 = 216 for the original developers, 12 + 48 = 60 for the new developer, and a total of 276 bugs. Second Talent test, 20 November 2026.

Test 5: A Job Post That Kept Every Rule

Length: 49 wordsBullets: 3Verdict: pass, best of the day

This was Dola's best answer. The post ran 49 words by our count, had three requirement bullets as asked, no emojis, and ended on the exact line "Apply via Second Talent.", punctuation included. Nothing came before or after the post itself.

Screenshot of Dola's job post for a Senior React Developer (Remote, Vietnam): a salary line, one sentence, exactly three requirement bullets and the final line Apply via Second Talent., with nothing before or after. Second Talent test, 20 November 2026.

None of the other three managed all of that. GLM-5.3 wrapped its post in an intro line and a closing question, MiMo dropped the closing period, and Perplexity put a dash in its heading.

Dola's one slip: it wrote the salary as a dollar range without the "USD" the prompt used.

Test 6: JSON Only

Search: 4 keywords, 24 sourcesVerdict: pass

The reply was a bare JSON array with no commentary and no code fence, and every value was right: Hanoi, VND, +07:00; Manila, PHP, +08:00; Jakarta, IDR, +07:00; Kuala Lumpur, MYR, +08:00.

Dola ran a web search to get there, 24 sources for facts that have not changed in decades, which cost a few seconds but did no harm.

Screenshot of Dola's JSON-only answer: an array of objects for Vietnam, the Philippines, Indonesia and Malaysia with capitals Hanoi, Manila, Jakarta and Kuala Lumpur, currency codes VND, PHP, IDR and MYR, and UTC offsets +07:00 and +08:00, with no commentary.

Test 7: Findings for a Paper That Does Not Exist

Search: 6 keywords, 36 sourcesVerdict: fail

Dola admitted it could not find the paper, then offered a summary of what such a paper would report, built on the "lattice-structured routing paradigm".

Three bold bullets followed: lattice routing improves load balancing, keeps accuracy at equal compute, and cuts routing cost. A closing note said these points "reflect the core empirical findings" of a "lattice routing family" of methods.

We found no such established line of work.

Screenshot of Dola asked to summarise a 2024 paper that does not exist: after Searched for 6 keywords, 36 sources it says it could not locate the paper, then gives three bold bullet findings on lattice-structured routing and a note calling them the core empirical findings of a lattice routing family of methods.

The disclaimer sits at the top, but the bullets read like a real summary, and anyone pasting them into a report would be quoting a paper that was never written.

Perplexity and GLM-5.3 both refused, named the closest real papers and asked for a link. For literature work, this is the result that rules Dola out.

What Is Dola? ByteDance's Doubao for the Rest of the World

Dola is the overseas version of ByteDance's Doubao assistant. Doubao serves mainland China; Dola serves other markets, under a different company and different laws.

The Star put it in one line in February 2026: Doubao's app "is called Dola, previously Cici, overseas". Both ciciai.com and cici.com now redirect to dola.com/chat.

Dola's terms of service, last updated 16 November 2026, name SPRING (SG) Pte. Ltd., a Singapore company, as the provider. The same name sells the app on the App Store and Google Play.

Chinese outlet Jiemian reported on 31 December 2025, citing people familiar with the matter, that Dola had passed 10 million daily active users. ByteDance has published no later figure.

Our Dola vs Doubao comparison sets the two apps side by side.

Jan 2024
Cici AI is one of four ByteDance apps for users outside China that Forbes reports running on OpenAI's GPT
Late 2025
Cici becomes Dola; the old domains now redirect to dola.com
Dec 31, 2025
10 million daily users, per Jiemian's sources
Mar 2026
BytePlus sells Seed models under the Dola name, starting with Dola Seed 2.0 Pro
Nov 16, 2026
iOS version 15.8.0 ships, and the terms of service are updated
Nov 17, 2026
Privacy policy updated, still naming SPRING (SG) as data controller

Where Dola Works, and Where It Does Not

Dola is the ByteDance assistant you can reach from outside China; Doubao turns most foreign visitors away.

On 17 November, doubao.com/chat sent our Southeast Asia connection to a region-ban page that offered two buttons: use Dola, or log in.

Screenshot of doubao.com's region-ban page reached from Southeast Asia on 17 November 2026, in Chinese, with an English translation below: region restricted, log in to use Doubao, or choose to use Dola, and two buttons, Use Dola and Log in.

Dola has gaps of its own. We queried Apple's store records for 19 storefronts on 20 November.

Apple lists the iPhone app in 15 of them, including Singapore, the Philippines, Vietnam, Indonesia, Malaysia, Thailand, Japan, South Korea, the UK, Germany, France, Brazil and Australia.

It is missing in the US, Canada, India and mainland China.

Google Play returned no Dola page for a US request either. We did not test dola.com from the US, so we cannot say whether the web chat opens there. Our list of Chinese AI chatbots notes which others work without a login from abroad.

Apps: Web, Desktop and Mobile

Dola runs in the browser, on desktop and on phones, and it pushes the desktop app hard.

The web chat shows a "Download for Windows" button at all times, and our first click as a guest opened a full-screen card headed "Access all features on the desktop app".

Screenshot of the full-screen card Dola shows a guest on the web: Access all features on the desktop app, a desktop AI assistant for work, study and daily tasks, with a Download Desktop App button over the dimmed chat page.
  • Web: dola.com/chat, guest chat with no account. Logging in offers Google, Facebook, Apple, a phone number or a QR code scanned with the phone app.
  • Desktop: the download page lists Windows x64 and ARM64 builds, plus Mac builds for Apple silicon, Intel and a Universal file. It pitches AI Docs, Slides and Sheets and scheduled research tasks.
  • iPhone and iPad: "Dola: Smart AI Assistant", version 15.8.0 from 16 November 2026, needs iOS 15 or later and rates 4.6 from 39,817 UK ratings.
  • Android: the Google Play listing shows 100M+ downloads and in-app purchases.

The chat window offers two modes, Chat and Work, plus shortcuts for Create Videos, Create Images, Writing, Homework and Translate.

Work is the agent mode that plans and runs multi-step tasks, but Dola's help centre, updated 19 November, says it is in beta, "not yet available to all users", with applications closed. We did not test it.

Is Dola Free? Plans and Prices

Yes, for chat. The App Store description says free users get basic chat, web search, chat about images and web pages, and a limited number of daily uses of work tools, image and video generation and voice calls.

Dola does not publish those daily limits.

Three paid plans, Plus, Pro and Max, raise the limits on "Pro mode", advanced image and video generation, spreadsheet tasks and research.

Dola has no public pricing page: its help article on plans links to a subscription page that asks you to log in first. The store listings carry the prices.

£19.90
Dola Plus, monthly (£199 a year)
£49.90
Dola Pro, monthly
£109.99
Dola Max, monthly
£14.90
Image and video pack, one-off
Source: in-app purchases on the UK App Store, read 20 November 2026. Singapore prices: S$29.98, S$69.98 and S$189.98 a month, pack S$19.98.

One detail matters for teams already on ByteDance's office suite: the help centre says a Lark account registered the same way shares a Dola plan's quota, and you can spend it in either app.

Apple's billing rules apply to plans bought on iOS.

Which Model Powers Dola?

ByteDance does not say. The chat window, the terms and the store listings name no chat model. The evidence points in two directions.

Toward ByteDance's own Seed models: BytePlus, ByteDance's cloud outside China, sells them under the Dola name.

Its model release notes describe dola-seed-2-1-turbo as "a next-generation model tailored for the coding and agent era", and the Seed 2.1 launch page has a "Try Dola" button that opens dola.com/chat.

Our guide to every Doubao model covers the China side.

Toward outside models: section 8 of the terms covers "Third Party LLMs", and the privacy policy says Dola may share chatbot content with "third-party developers of integrated AI tools, including LLMs such as Gemini (Google)".

Cici ran on OpenAI's GPT when it launched. Our reading: Seed models most likely answer most chats, but Dola keeps the right to route some to partners, and you cannot tell which one replied.

Image and Video Generation

The AI Creation page names its models, unlike the chat. For images, Seedream 4.5 is the default, "reliable quality for daily use", and Seedream 5.0 Pro costs four times the credits.

For video, the default is Dreamina Seedance 2.0 Fast with a 10-second clip, and Seedance 1.0 is the option "for simple videos".

Screenshot of Dola's AI Creation page with the image model menu open: Seedream 5.0 Pro, labelled professional use and 4x credit usage, and Seedream 4.5, labelled reliable quality for daily use, above a grid of sample images. Captured 20 November 2026.

The app trails ByteDance's own catalogue here. BytePlus released Dola Seedream 5.5 to developers in October 2026, with layered RGBA output, but the consumer app still topped out at 5.0 Pro on 20 November.

The desktop download page sells the pair for posters, comics, storyboards and e-commerce ads.

We did not test image or video output in this round; our roundup of Chinese AI video generators covers Seedance alongside its rivals.

Privacy and Where Your Data Goes

The privacy policy, updated 17 November 2026, names SPRING (SG) Pte. Ltd. as data controller. It collects prompts, chat history, uploaded files, voice recordings and photos, plus your IP address and an approximate location derived from it.

It also says Dola may infer "gender, age and hobbies" from what it has.

  • Storage: servers in Malaysia, Singapore and Japan, per the section for Europe and the UK; the Brazil and Latin America sections name Malaysia and Singapore.
  • Model training: your inputs can train AI models, including models used across ByteDance's corporate group. The opt-out is a toggle in account settings, which a guest does not have.
  • Sharing: with ByteDance group companies, advertising and measurement partners, and third-party LLM providers such as Google.
  • Age: 13 or older to open an account, higher where local law says so.

The policy you see depends on where you are. The footer link redirects by region, and from Vietnam it served a Vietnamese version effective 17 December 2025.

The terms add two points for business users: you own your input and Dola claims no ownership of output, but the service is "provided for private and non-commercial use", under Singapore law with arbitration at the SIAC.

Dola vs ChatGPT, Qwen and Perplexity

On our seven tests, Dola finished level with MiMo and behind Perplexity and GLM-5.3. Its two outright fails were the tests that punish a confident wrong answer: debugging and the fake paper.

Heatmap of Second Talent's seven hands-on tests on 20 November 2026. Dola: partial on the password checker, fail on the sort bug and the fake paper, pass on the fact-check, maths, job post and JSON. Perplexity fails only the maths. GLM-5.3 passes all seven. MiMo-V2.6-Pro fails the fact-check and is partial on the sort bug, job post and fake paper.

GLM-5.3 ran on a free Z.ai account at Max effort, Perplexity on its free plan, MiMo-V2.6-Pro signed in to MiMo Studio, and Dola as a guest.

ChatGPT and Qwen did not get the full battery; the table uses their results from our 17 and 20 November rounds.

DolaChatGPTQwen ChatPerplexity
MakerByteDance (SPRING SG)OpenAIAlibabaPerplexity
Chat without an accountYesNot testedLog-in box on send (20 Nov)No, sign-in after the first prompt
Cheapest paid plan£19.90 a month (UK)Go, $8 a monthNo public USD price foundPro, $17 a month billed yearly
Debugging testWrong (17 and 20 Nov)Correct, less precise (17 Nov)Correct (17 Nov)Correct (20 Nov)
Fact-check testCorrect, 3 linksNot runCorrect, 16 sources (20 Nov)Correct, 5 sources (20 Nov)
Fake-paper trapInvented findingsNot runNot runRefused, asked for a link

Prices are from ChatGPT's pricing docs (Plus is $20) and Perplexity's pricing page, both read on 20 November 2026. Perplexity is the closer rival for sourced answers, and our Perplexity guide has its full results.

For coding, Claude and ChatGPT both found the sort bug Dola missed, and so did Qwen.

Who Should Use Dola: Our Verdict

Dola is a fast, free, no-login assistant for everyday questions, and an unsafe one for anything you cannot check. Its wrong answers read with the same certainty as its right ones, which is the problem with the sort bug and the fake paper.

Use Dola for
  • Quick factual questions with links you can open
  • Short writing with strict rules, like job posts
  • Structured output such as JSON and simple maths
  • Images and short videos on the Seedream and Seedance models
Skip Dola for
  • Code review and debugging: failed twice
  • Summaries of papers or reports you have not read
  • Company work: the terms say non-commercial use
  • Users in the US, Canada or India: no App Store listing

If you need a free assistant outside China and value speed over depth, Dola is worth a bookmark. If you write code or research for a living, Perplexity or GLM-5.3 did more of our tests right on the same day.

Hire Engineers Who Check What AI Writes

Dola's wrong debugging answer came back in under 30 seconds and read as certain, which is why AI output in production still needs an engineer who reviews it.

Second Talent places vetted LLM developers from Asia who build on Seed, Qwen and DeepSeek. Tell us who you need and we will send a shortlist within 24 hours.

Frequently Asked Questions

Is Dola the same as Doubao?

No. They are separate apps from ByteDance with separate accounts and operators. Doubao serves mainland China; Dola, run by SPRING (SG) in Singapore, serves other markets.

Is Dola safe to use?

It is a mainstream app from a large company, but read the policy first. Chats can train ByteDance's models unless you opt out in account settings, and some content can go to third-party models such as Gemini.

Keep passwords and confidential data out of it, as the policy itself asks.

Can I use Dola without an account?

Yes. In our 20 November test, guest chat worked with no login. On 17 November it asked us to confirm we were at least 18 first.

Does Dola work in the US?

The iPhone app is not listed on the US App Store, and Google Play showed no page for a US request on 20 November. We did not test the web chat from the US.

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

Start Hiring $0 cost until you hire
Matt Li

Written by

Matt Li is a tech-driven entrepreneur with deep expertise in global talent strategy, digital experience optimization, e-commerce, and Web3 innovation. He is the Co-Founder of Second Talent, a US-based company that connects businesses with top-tier tech professionals worldwide. Since launching the company in 2024, Matt has led its growth by leveraging technology to streamline remote hiring and scale distributed teams. With a background spanning product, operations, and innovation, Matt brings a cross-disciplinary perspective to the evolving digital economy. His work sits at the intersection of global talent, emerging technology, and scalable digital transformation.

More posts by Matt Li →