Chart of the Day · SEP 22, 2026

Ollama Downloads: Is Local LLM Adoption Keeping Up?

Ollama's JavaScript client grew 2.1×, but the cloud SDKs grew 8.7×. Local AI is growing; it is not where most products are built.

Mode VC ·

The signal

On npm, Ollama's client fell from 27 to 6.6 downloads per 1,000 cloud lab SDK downloads in 46 weeks. On PyPI, after its counting change, it is 21.5.

Weekly downloads of the ollama client per 1,000 combined downloads of the OpenAI, Anthropic and Google SDKs. npm: complete weeks, 6 Oct 2025 – 30 Aug 2026. PyPI: complete weeks after the 25 Aug 2026 counting change only.

npm (JavaScript)PyPI (Python), after 25 Aug
Sources: npm registry downloads API · pypistats.org API (post-25 Aug 2026 only) · Official launch posts (launch feed, captured 7 Oct 2026)MODE VC

The ollama package is one client for a local model server. Apps can also reach local models through other clients, and local tools that are not on npm or PyPI are not counted, so this is a lower bound on local use. Downloads are installs (including CI), not users. PyPI changed how it counts on 25 Aug 2026; we use only whole weeks after that date and only ratios, never PyPI levels before and after. Weeks with an npm gap day are dropped.

Show the data (46 rows)
RegistryWeek ofollamaOpenAI + Anthropic + Google SDKsper 1,000
npm6 Oct 20250.32M11.7M27.2
npm13 Oct 20250.31M12.2M25.8
npm20 Oct 20250.32M11.7M27.1
npm27 Oct 20250.30M11.6M25.7
npm3 Nov 20250.29M12.2M23.5
npm10 Nov 20250.31M12.3M25.3
npm17 Nov 20250.43M13.3M32.4
npm24 Nov 20250.45M12.1M37.3
npm1 Dec 20250.53M14.4M36.8
npm8 Dec 20250.40M14.5M27.7
npm15 Dec 20250.29M14.4M20.3
npm22 Dec 20250.23M9.4M24.6
npm29 Dec 20250.24M9.6M24.9
npm5 Jan 20260.34M15.7M21.9
npm12 Jan 20260.35M16.7M20.8
npm19 Jan 20260.44M18.2M24.1
npm26 Jan 20260.60M21.8M27.6
npm2 Feb 20260.67M22.8M29.5
npm9 Feb 20260.59M23.9M24.8
npm16 Feb 20260.49M23.4M21
npm23 Feb 20260.43M28.8M14.9
npm2 Mar 20260.46M31.4M14.6
npm9 Mar 20260.48M35.4M13.5
npm16 Mar 20260.43M34.9M12.4
npm23 Mar 20260.47M35.9M13.1
npm30 Mar 20260.44M37.1M11.9
npm6 Apr 20260.46M40.1M11.4
npm13 Apr 20260.46M42.9M10.7
npm20 Apr 20260.47M45.3M10.4
npm27 Apr 20260.50M50.7M9.9
npm4 May 20260.56M48.0M11.6
npm11 May 20260.61M50.4M12
npm18 May 20260.58M54.0M10.7
npm25 May 20260.52M61.9M8.5
npm8 Jun 20260.54M65.2M8.3
npm15 Jun 20260.55M68.8M7.9
npm29 Jun 20260.52M63.7M8.1
npm13 Jul 20260.65M72.7M9
npm20 Jul 20260.65M74.7M8.7
npm27 Jul 20260.71M80.2M8.9
npm3 Aug 20260.66M79.3M8.3
npm17 Aug 20260.63M89.6M7
npm24 Aug 20260.67M101.8M6.6
PyPI31 Aug 20262.8M125.3M22.1
PyPI7 Sep 20262.8M128.6M21.8
PyPI14 Sep 20262.9M133.1M21.5

The Mode read

Our launch feed confirms open weights for 1 of the 44 models launched since 6 July; 14 are closed and the rest are unconfirmed. Open weights are what local AI runs on. So how much are developers building on local models compared with cloud APIs?

Less and less, relative to the cloud. On npm, Ollama's client grew 2.1×, but the OpenAI, Anthropic and Google SDKs together grew 8.7×. Its rate fell from 27 to 6.6 installs per 1,000 cloud SDK installs, 4.1× lower. Python is different: in the week of 14 Sep, after PyPI's counting change, the rate there was 21.5, 3.3× the latest JavaScript rate.

JavaScript is where products ship to users, and there the cloud is winning by a wide and widening margin. Local models show up more in Python, the language of experiments, fine-tuning and data jobs. For founders, local AI is not a mass developer market. It is a sales motion: buyers who cannot send data to a cloud API, reached through the vendors and IT channels they already trust.

Local models win on privacy. The cloud wins on distribution, and it is pulling away.

What we would back: Companies that package open-weight models for buyers who must keep data in-house, sold through channels those buyers already use.