Business judgment on AI products
DeepSeek's phone assistant: the same model, for search, writing, reading, problem-solving, and translation on the go.
01
Start inside the user's day · Public facts + observable behavior · 2026-09-01
Users need to complete tasks such as search, writing, reading, problem-solving, and translation, and seek to improve efficiency with an AI assistant.
Currently users may use other AI assistants, search engines, or manual work.
These tasks are time-consuming and require expertise; manual completion is inefficient and costly.
The trend is one model split into a website and a phone app, with the phone taking scraps of time. The entry is not another general chat — homework lookup, translation, and commute writing. Consumer use is free; money still sits on the API.
Public records show 139.08M MAU, indicating significant attention and usage, but retention and payment are not yet verified.
① Whether downloads keep recovering for two consecutive quarters — if not, the app; line's decline is structural; ② Whether MAU turns YoY-positive in H2 2026 — it is currently the only shrinking; top-tier product; ③ Any new "explosion moment" (the traffic pulse of a model release) — DeepSeek's; user…
Worth dissecting. Public records show 139.08M MAU, indicating significant attention and usage, but retention and payment are not yet verified.
if you build a "tool" product (users come, use, leave), copy the entry split — deep work lives on the web, the app is a mobile supplement, and do not expect the app alone to pull general users. Tool products grow on "highlight moments" (model releases, viral events), not on daily compounding; allocate budget accordingly. the free-consumer-plus-metered-B-end double layer transfers — losing money on the consumer side is fine if brand and ecosystem value feed back into the API/B-end. It only works if your consumer product has strong brand momentum; otherwise free is just giving it away.
End users: free, no ads; Developers: token-metered API, long-term among the cheapest domestic models; Profit center: not disclosed separately; open-weights plus a free product is the; acquisition move, API and B-end are the monetization cha… ① Whether downloads keep recovering for two consecutive quarters — if not, the app; line's decline is structural; ② Whether MAU turns YoY-positive in H2 2026 — it is currently the only shrinking; top-tier product; ③ Any new "explosion moment" (the traffic pulse of a model release) — DeepSeek's; user…
DeepSeek's phone assistant: the same model, for search, writing, reading, problem-solving, and translation on the go.
Public records show 139.08M MAU, indicating significant attention and usage, but retention and payment are not yet verified.
An English validation note will follow from the public evidence.
The product claims to help users complete: “DeepSeek's phone assistant: the same model, for search, writing, reading, problem-solving, and trans”. User evidence has not yet verified pain intensity or the cost of doing without it.
The assessment is recorded; an English explanation is pending.
The assessment is recorded; an English explanation is pending.
The assessment is recorded; an English explanation is pending.
02
Market comparison
The Chinese–English market comparison is not complete yet. A conclusion follows only after its coverage and verifiable evidence are recorded.
03
The call and next move come first; the full read retains the evidence and counterevidence.
DeepSeek's mobile app — the second entry point for the same model and the same brand: search, writing, reading, problem solving and translation packed into the phone, catching the fragmented scenarios the web product does not cover.
DeepSeek, the Chinese AI company known for open-weights models and low-cost training. This slug maps to the mobile app entry; the web product (chat.deepseek.com, slug: deepseek) shares the same model family. The app and the web product are not a main-site/affiliate relationship but two entries of one product line on phone and PC: the web is for deep use (long documents, code, research), the app is for mobile Q&A and utility scenarios.
Read: splitting one model into a web product and an app is standard practice for Chinese platforms (the pool's inspiration note says as much). DeepSeek does it to cover two kinds of people — heavy users who stay on the web and light users who come in from the phone. But running a consumer app has never been a model company's strength — the data below proves the point.
What it deliberately does not do: no entertainment feeds, no short video, no AI companionship. QuestMobile classifies it in the "AI search engine" track — the only search-class AI app in the top three of the entire chart. Its product personality is "a tool you do work with," not "a toy to kill time."
It replaces the everyday "general search + Q&A" routine, but aimed at a more professional crowd. Asking "what does this paper say," "why does this code error," "how is this function derived" used to mean opening a search engine and digging through links, or separately opening a translator, a formula editor, a forum. The DeepSeek app collapses it into one action: paste or photograph, get the answer, keep asking. This replacement is not for everyone — QuestMobile and media analysis both find its new users are concentrated among programmers, researchers and independent media workers. It replaces "deep professional scenarios," not "quick weather checks."
For heavy users, it replaces parts of the old office suite: writing an industry report used to mean Word, a dictionary for translation, a calculator for math. One dialog box now covers them. That replacement is real but the crowd is limited.
Read: DeepSeek's consumer-side business model barely exists — the app is free with no ads and does not make money. Its job is brand entry and model distribution; the money comes from API and enterprise. That is why its app investment will stay limited: not because it cannot be done well, but because there is no need to monetize the app. That is the biggest structural difference from Doubao — Doubao has to live off C-end users; DeepSeek does not.
| Dimension | Call |
|---|---|
| Founder-product fit | A model company building an app — medium fit: strong model, weak consumer operations |
| Product insight | No distinctive interaction innovation; wins on the model's first-try correctness |
| Execution quality | In-house MoE, ultra-long context, open ecosystem — top-tier domestic technical base |
| Timing | The window passed: the early-2025 explosion brought users, but general users do not stay |
Stable leader, but the app line is shedding users — proof that two legs walk differently.
App MAU of 139M (traffic board) or 130M (QuestMobile, June) is first tier among domestic AI apps, and the numbers are real. But the trend is negative: -20.3% YoY, four quarters of download decline, DAU an order of magnitude below Doubao. Nothing is broken with the product; it is the natural boundary of "free tool plus deep scenario." The pool of programmers and researchers has a ceiling, and general users come and leave because there is no reason they must use DeepSeek — Doubao does the entertainment, companionship and quick-answer layers much heavier.
The -1.5% MoM matches the web product's -10.59% (slug: deepseek) — the shrink is shared across both entries, not an app-only problem. DeepSeek's signature has never been "retaining users"; it is "open-weights plus free pricing that pushes the whole industry's price to the floor." That shock keeps spreading on the B-end and in the ecosystem; app-user fluctuation barely touches its fundamentals.
Entry relationship: the web is the home of deep use, the app is the mobile supplement. Neither is an affiliate/overseas version — they share the brand and models, and split the time of the same users.
① Whether downloads keep recovering for two consecutive quarters — if not, the app line's decline is structural ② Whether MAU turns YoY-positive in H2 2026 — it is currently the only shrinking top-tier product ③ Any new "explosion moment" (the traffic pulse of a model release) — DeepSeek's user numbers have always been tied to model launches
Product logic: if you build a "tool" product (users come, use, leave), copy the entry split — deep work lives on the web, the app is a mobile supplement, and do not expect the app alone to pull general users. Tool products grow on "highlight moments" (model releases, viral events), not on daily compounding; allocate budget accordingly.
Pricing structure: the free-consumer-plus-metered-B-end double layer transfers — losing money on the consumer side is fine if brand and ecosystem value feed back into the API/B-end. It only works if your consumer product has strong brand momentum; otherwise free is just giving it away.
Worth watching. The top-tier position is real, but the app line is losing users with no reversal signal yet. It is evidence of the rule that "a model company's app operations are its weak spot" — and if downloads are still falling in six months, the conclusion gets stronger. The watch point is not the app itself but whether the next model release reignites the traffic pulse.
04
Evidence trail
05
Use these searches when the official site is missing or the current link is only a lead.