Developer counts: why "23.33M" and "500k" are not comparable
When different companies quote "developer numbers" they are not talking about the same thing. Baidu cites 23.33M developers in the Paddle-ERNIE ecosystem; DeepSeek has roughly 500k. Comparing those two directly is like comparing "developers across an entire OS ecosystem" with "users of a single app".
This page addresses one problem: the same sentence, three denominators.
Framework ecosystem developers, model developers and paying API customers are three different
funnel layers. Mixing them produces completely wrong conclusions — and that mistake is common.
1. Each company's layers
ERNIE / PaddlePaddle (Baidu)
| Layer | Value | Basis | Source | Confidence |
|---|---|---|---|---|
| ① Framework ecosystem developers | 23.33M | Cumulative registrations under PaddlePaddle (academia, enterprise, individuals) | Baidu CTO, Sep-Nov 2025 | Medium |
| ② Qianfan platform enterprise customers | 460k | Cumulative enterprises served by the Qianfan platform | Baidu official, Oct 2025-Feb 2026 | Medium |
| ③ Agents built on the platform | 1.3M | Cumulative agents built on Qianfan | Baidu official | Medium |
| ④ ERNIE model API customers | 85k | Customers calling the ERNIE API directly | ERNIE 4.0 disclosure, Apr 2024 | Low |
Baidu counts developers across its whole stack, starting from the framework layer.
DeepSeek
| Layer | Value | Basis | Source | Confidence |
|---|---|---|---|---|
| ① Framework ecosystem developers | — | No proprietary framework ecosystem (uses PyTorch et al.) | — | — |
| ② Platform enterprise customers | — | No official platform figure | — | — |
| ③ Agents built on the platform | — | 8,559 Harness plugins (what we index) | Our index | High |
| ④ Registered developers (estimate) | ~500k | Model platforms (HF, ModelScope) | Yanjingbizhi, Apr 2026 (relayed, low confidence) | Low |
| ⑤ Open-source contributors | ~12k | Actually contributing, not just downloading | Same report | Low |
| ⑥ Paying API customers | ~41k | Paying directly on the open platform | presenc.ai (estimate) | Low |
| ⑦ On-prem deployments | ~11.4k | Running DeepSeek models on-prem or in private cloud | presenc.ai (estimate) | Low |
DeepSeek counts people building on its models. There is no framework layer.
Qwen (Alibaba)
| Layer | Value | Basis | Source | Confidence |
|---|---|---|---|---|
| ① Derivative models | Over 200k | Community models fine-tuned from Qwen weights | Alibaba official | Medium |
| ② Global downloads | Over 1B | Cumulative downloads on HF and similar | Alibaba official | Medium |
| ③ Languages supported | 119 | Model capability basis | Alibaba official | Medium |
Qwen uses derivative count and downloads as proxies for developer scale.
2. Only comparable once aligned
| Comparison | A | B | Verdict |
|---|---|---|---|
| Comparable layer: paying API customers | ERNIE 85k (Apr 2024) | DeepSeek ~41k (estimate) | Same order of magnitude |
| Not comparable: framework ecosystem vs model developers | Paddle 23.33M | DeepSeek 500k (estimate) | 46x apart, but the denominators are unrelated |
Key conclusion: Baidu's 23.33M counts cumulative developers in the
PaddlePaddle framework ecosystem, not developers using the ERNIE model. The figure genuinely
comparable to DeepSeek is ERNIE's 85k API customers — and those two are the same order of
magnitude.
3. Why ERNIE "looks like it has no developers"
| Reason | Detail |
|---|---|
| Low open-source visibility | In 2024 Baidu's CEO publicly called open-source models "an IQ tax". ERNIE stayed closed until the 4.5 series moved to a dual track (open base, closed pro). |
| Confusing consumer metrics | The app has been renamed repeatedly (ERNIE Bot to Wenxiaoyan to ERNIE), so Baidu's "200M MAU" and QuestMobile's 5.31M for Q3 2025 are not measuring the same thing at all. |
| Fragmented developers | Paddle, Qianfan, AI Studio and the mentor programme are all counted separately, so outsiders cannot assemble one picture. |
4. Aside: why plugins are "only half tested"
From our own survey of the 24 highest-starred DSH plugins:
Have CI
100%
Have unit tests
46%
Have integration tests
~0%
It is not that nobody tests — it is that only half the job is done: functions are tested, but "will it install into a real DSH" is not.
| Reason | Detail |
|---|---|
| Cost: a unit test takes seconds, an integration test takes minutes | Integration needs DSH installed, the plugin installed, a boot, and log capture. Complex setup and expensive CI minutes. |
| Incentive: authors are only accountable for their own environment | Individual developers ship once it works on their own macOS or Windows. No QA team, and no perceived need to cover a Linux matrix. |
| Technical: needs a real DSH instance | You cannot test functions alone — you must install into a real DSH and see it load. Plugins can also conflict, so the environment must be clean. |
| No third party requires it | No platform enforces "must install successfully" before listing, so there is no pressure. |
That gap is our position. Integration testing costs fall on the
author while the benefit is public — so nobody does it. We absorb the cost and publish the results
as a public good. That is where a testing platform belongs in this ecosystem.
5. Measurement caveats
| Caveat |
|---|
| Framework ecosystem developers, model developers and paying API customers are three funnel layers. Mixing them produces completely wrong conclusions. |
| Comparing Baidu's 23.33M to DeepSeek's 500k is like comparing "developers across an entire OS ecosystem" to "users of one app". |
| The ERNIE app has been renamed repeatedly (ERNIE Bot to Wenxiaoyan to ERNIE), so third-party MAU trackers disagree wildly. |
| The 85k ERNIE API customer figure dates from Apr 2024 and is stale — but it is the only figure comparable to DeepSeek's 41k. |
| The 23.33M figure was last published in Sep-Nov 2025; no 2026 update yet. |
6. Sources
| Source | Data | Type |
|---|---|---|
| 百度 Create 大会 / 王海峰 | Baidu Create / CTO Wang Haifeng: 21.85M Paddle-ERNIE developers (Apr 2025) | 官方披露 / official |
| 百度官方（2025-09—11） | Baidu official (Sep-Nov 2025): 23.33M developers, 760k enterprises, 1.1M+ models, 930k commits, 18k contributors | 官方披露 / official |
| 千帆平台（2025-10—2026-02） | Qianfan (Oct 2025-Feb 2026): 460k enterprises, 1.3M agents, 150+ models | 官方披露 / official |
| 文心 4.0（2024-04） | ERNIE 4.0 (Apr 2024): 85k API customers, 200M daily calls | 官方披露（较旧）/ official (older) |
| 研精毕智（2026-04） | Yanjingbizhi (Apr 2026): DeepSeek ~500k registered, 12k contributors | 转引，可信度低 / relayed, low confidence |
| presenc.ai | presenc.ai: DeepSeek 41k API customers, 11.4k on-prem | 估算 / estimate |
| QuestMobile | QuestMobile: Wenxiaoyan 5.31M MAU (Q3 2025) vs Baidu's "200M MAU" | 第三方 vs 公司口径冲突 / conflicting bases |
Our stance: we are not ranking ecosystems, only clarifying
denominators. A number's size depends on how you count, not on who is stronger.
Related
User and developer scale · Three-way comparison · Plugin ecosystem