How many users and developers does DeepSeek actually have
About 130M monthly active in China, around 139M globally — but these numbers badly undercount real usage. Two reasons: most users run DeepSeek inside other products without knowing it, and the developer count has never been published officially.
Our read: C-end growth has stalled while B-end is exploding — real usage
is going invisible. Someone using the AI in a note app, a support desk or a code-completion
plugin may be running DeepSeek underneath, but neither QuestMobile nor the App Store counts them as a
"DeepSeek user". The 100x growth in API token volume is the sum of exactly those indirect users.
1. C-end and B-end metrics
| Metric | Value | Basis | Source | Confidence |
|---|---|---|---|---|
| China MAU | 130M (May 2026) | China app | QuestMobile H1 2026 report / TMTPost | Medium |
| China ranking | No.3 (Doubao 382M, Qwen 167M) | China app | QuestMobile | Medium |
| China YoY | -20.3% (only one of the top three shrinking) | China app | QuestMobile | Medium |
| Global MAU | ~139M (Apr 2026 estimate) | Global app | AICPB (via Thunderbit) | Low |
| Cumulative downloads | ~173M (since Jan 2025) | Global app downloads | AppMagic (via Business of Apps) | Medium |
| Quarterly downloads | 83M (Q1 2025) to 13M (Q1 2026) | Global app downloads | AppMagic | Medium |
| Website visits | 354M/month (Aug 2026) | Web traffic | Similarweb | Med-high |
| Relative web scale | About 6% of ChatGPT's website | Web traffic | Similarweb | Med-high |
| Historic peak | Topped App Store free charts in 157 countries (Jan 2025) | App Store charts | Public reports | Medium |
| Global AI app download rank | No.3 (behind ChatGPT and Gemini) | Global app downloads | Sensor Tower, State of Mobile 2026 | Med-high |
| ARR (annualised run rate) | USD 1B (Sep 2026) | Last month x 12 projection, not cash received | The Information, 24 Sep 2026 | Medium |
| Recognised revenue | RMB 475M (Jan-Jul 2026) | Cash received | The Information / Sacra | Medium |
| 2025 full-year revenue | ~USD 7M | Cash received | The Information | Medium |
| API gross margin | 82.9% (first 7 months of 2026) | API segment | The Information | Medium |
| Blended gross margin | 44.6% | All segments | The Information | Medium |
| Net loss | ~RMB 715M (first 7 months of 2026) | All segments | The Information | Medium |
| API customers | ~41k (Q1 2026) | Estimate | presenc.ai | Low |
| On-prem deployments | ~11,400 companies | Estimate | presenc.ai | Low |
| Enterprise signings (Chinese source) | V4-Pro over 12k (Apr 2026) | Second-hand | Yanjingbizhi report | Low |
| OpenRouter weekly request share | 26.5% (week of 19 Oct 2026, No.1) | Publicly verifiable | OpenRouter | High |
| OpenRouter token share | 9% (Jan) to ~18% (Jun) | Publicly verifiable | aboutchromebooks / Sacra | High |
| Weekly token volume | 0.31T (Jan) to 35.4T (Oct) | Publicly verifiable | OpenRouter | High |
| V4.1-Flash-1017 single week | 33.6T tokens (+801% WoW) | Publicly verifiable | OpenRouter | High |
| Harness open-source repo | 260k stars (Nov 2026) | Verifiable on GitHub | GitHub | High |
Two rows matter most: China MAU is down 20.3% year over year,
the only one of the top three shrinking; and ARR of USD 1B sits about 14x above recognised
revenue of RMB 475M. The C-end has peaked while the B-end explodes.
2. Developer count: never published, four conflicting denominators
This is the part that needs the most care. Every figure in circulation is a third-party estimate, and they contradict each other.
| Basis | Value | Scope | Source | Confidence |
|---|---|---|---|---|
| Registered developers | ~500k | Model platforms (HF, ModelScope) Relayed by self-media; mixed vintage, weak sourcing | Yanjingbizhi report, Apr 2026 | Low |
| Open-source contributors | ~12k | Actually contributing, not just downloading Far smaller than registrations; closer to real active developers | Same report | Low |
| Paying API customers | ~41k | Paying directly on the open platform Third-party estimate, not official | presenc.ai (estimate) | Low |
| Enterprise API users | 300k+ | May include anyone routed via third-party clouds Seven times the 41k figure because the denominator differs | Yanjingbizhi report | Low |
Why the spread. Registered is not active, and active is not paying:
500k registered, 12k contributors, 41k paying API customers are three different funnel layers.
The question "how many developers" is unanswerable until you say which layer.
And "300k enterprise users" may count anyone routed through a third-party cloud, so it shares no
denominator with "41k direct payers" — they cannot be added or compared.
3. Verifiable proxy metrics
Platform data is harder to fake than headcount estimates:
| Metric | Value | Source | Confidence |
|---|---|---|---|
| HuggingFace cumulative downloads | 75M (Feb 2026) | NetEase Finance (relayed) | Medium |
| V4-Flash downloads | 5M (Apr 2026) | Yanjingbizhi | Low |
| Harness launch day | 84,689 stars / 7,482 forks / 12,293 commits | GitHub | High |
| Harness beta phase | 769 developers, 712 repos | GitHub | High |
| OpenRouter weekly request share | 26.5% (No.1) | OpenRouter | High |
4. Where C-end users are
The three columns use completely different denominators. Do not compare them.
| Country | Download share | MAU share | Web share |
|---|---|---|---|
| 中国 / China | 32% | 30.71% | 51.43% |
| 俄罗斯 / Russia | 9% | — | 8.27% |
| 印度 / India | 6% | 13.59% | — |
| 美国 / US | 5% | 4.34% | 5.86% |
| 印尼 / Indonesia | 3% | 6.94% | — |
| 巴西 / Brazil | 3% | 3.13% | — |
| 巴基斯坦 / Pakistan | 3% | — | — |
| 法国 / France | 2% | 3.21% | — |
Download share: AppMagic full-year 2025 ｜ MAU: AICPB 2025-01 ｜ Web: Similarweb 2026-08
5. B-end by region (weekly active)
| Region | Share |
|---|---|
| 中国 / China | 47% |
| 东南亚 / SE Asia | 14% |
| 美国 / US | 11% |
| 印度 / India | 9% |
| 欧洲 / Europe | 8% |
| 拉美 / LatAm | 4% |
| 中东非 / MEA | 4% |
6. Pricing is tightening
B-end is moving from "subsidise for growth" to "charge for value". Four changes in four months:
| Date | Change | Detail |
|---|---|---|
| 2026-08-17 | Peak/off-peak pricing | V4-Pro cached input RMB 0.30 per 1M tokens (+1100% vs promo) |
| 2026-09-10 | Flash price cut | Cached input down to RMB 0.02 (~-60%) |
| 2026-10-19 | Search and crawl billed per request | Moved from per-token to per-request |
| 2026-11-01 | Web-search API charge | RMB 5 per 1,000 successful requests |
7. Why real usage is far above 130M
| Signal | Meaning |
|---|---|
| 26.5% of OpenRouter requests — number one | One in four developer API calls goes to DeepSeek |
| Weekly tokens up from 0.31T to 35.4T | Usage up 114x while MAU falls — the gap must be indirect use |
| Chinese models on Vercel: 11% to 55% of token traffic | Enterprise developers are migrating; end users cannot tell |
| Ramp: fastest-growing AI vendor in US B2B spend | B-end substitution is happening |
| 136.7 minutes per user per month — longest of the three | Remaining users engage more heavily, but the total shrinks |
In other words: DeepSeek is turning from an app into a layer of infrastructure. Infrastructure does not seek recognition. It seeks dependency.
8. Traps to avoid before quoting these numbers
| Trap | Detail |
|---|---|
| "USD 1B ARR" is not "RMB 475M revenue" | About 14x apart. ARR is a last-month-times-12 projection, not cash. Media mixes them. |
| 41k API customers and 11.4k deployments are estimates | From presenc.ai, not official disclosure. Use for trend, not precision. |
| "300k enterprise users" vs "41k API customers" differ 7x | The former may count anyone routed through third-party clouds; the latter only direct payers. |
| Geographic denominators do not match | Download share, weekly-active share and web-traffic share use three different denominators. |
| MAU estimates vary widely | In April 2025 the range was 38M to 96.9M depending on the tracker. |
| Website traffic counts only direct C-end use | B-end API calls never touch deepseek.com, so they are invisible in web traffic. |
| DeepSeek has never published an official developer count | Every figure in circulation is a third-party estimate, and they conflict. |
9. Sources
| Source | Used for | Type |
|---|---|---|
| QuestMobile | China MAU, ranking, YoY, profile | Third-party tracker |
| AppMagic / Appfigures / Sensor Tower | Downloads and ranking | Third-party tracker |
| Similarweb | Web traffic | Third-party tracker |
| AICPB / presenc.ai | Global MAU, API customers, deployments | Estimate |
| OpenRouter | Request share, token volume | Publicly verifiable |
| Sacra | Token trends, ARR breakdown | Research firm |
| The Information | ARR, revenue, margin, loss | Media (company figures) |
| Ramp | US B2B spend growth | Platform data |
| 研精毕智（Yanjingbizhi） | Developer counts, downloads | Relayed, low confidence |
Our stance: we compile public figures and label the basis.
We do not publish unsourced projections. Estimates are marked and never mixed with verifiable
data. If a number disagrees with the original report, tell us.
Related
DeepSeek vs Qwen vs Doubao · Company · Timeline · Official resources
Token volume trend
Weekly token volume for DeepSeek models on OpenRouter. These are published data points connected by a line — not a continuous series.
Latest week
35.4T
Since January
+11,319%
Per day
5.06T
Per second
58.5M
Within that, the DeepSeek Harness app alone processed 2,237,022,829,448 tokens through OpenRouter this week (rank 12) — that figure is from our own daily collection.
How to read it. The line connects published weekly figures.
We add a point only when a new weekly number appears, so a steep segment means "average growth
between two announcements", not a smooth daily climb. New points get appended as our daily
collection continues.
Live conversion
At the latest weekly rate, DeepSeek processes roughly 58,531,746 tokens per second. In the time it takes to read this paragraph (about 20 seconds), about 1,170,634,920 tokens flow through.
Since you opened this page
0
tokens (extrapolated from the weekly average, not a live count)
Sources
| Data point | Trillion/week | Source |
|---|---|---|
| 2026-01-05 | 0.31 | OpenRouter / Sacra |
| 2026-07-27 | 10.50 | OpenRouter / Sacra |
| 2026-08-03 | 17.30 | OpenRouter / Sacra |
| 2026-10-17 | 33.60 | OpenRouter / Sacra |
| 2026-10-26 | 35.40 | OpenRouter / Sacra |
OpenRouter does not expose model-level token volume via API. Figures above come
from its public leaderboard and weekly reports relayed by third parties. We also collect app-level
data daily; the series lives at /data/openrouter-tokens.json.