An update and licensing service for applications that ship on-device models: a CDN-backed delivery layer that streams quantised model weights to a desktop app in resumable chunks, selects the right variant for the detected hardware, ships deltas when a model is updated, and issues signed licence tokens so a commercial app can monetise a locally running model. The boring infrastructure that on-device AI products need and none of them are building.

Market Panorama

Competitive Intelligence

Users & Demand

Business Model

Market Sizing

Tech Solution & Build Plan

Go-to-Market Strategy

Risk Assessment

Action Recommendations

### Market definition and scope

The local model delivery service covers everything a desktop application needs between "we have quantised weights" and "our users are running them": chunked and resumable distribution, hardware capability detection that selects the right quantisation, content-defined delta updates so a model refresh does not cost three gigabytes, and signed licence tokens that enforce paid or seat-limited access for models sold as part of the product.

The boundary excludes three adjacent domains. It is not a runtime — Ollama and llama.cpp run models well and leave delivery to the application. It is not research hosting — Hugging Face serves weights to researchers and has no product-delivery semantics. And it is not model conversion, which is the step before delivery.

Its unit of value is a number a developer can put in a launch post: the update went from 2.8 GB to 47 MB and the download completion rate went from 71% to 96%.

October produced a cohort, not an anecdote. Survival LLM, an AI assistant and emergency library that runs entirely on local hardware, reached 543 points and 204 comments (Hacker News, October 2026). Hob, an open-source on-device alternative to Claude Cowork and ChatGPT, reached 179 points and 43 comments. Pulse Local, a Mac budgeting app whose models run on the machine, reached 122 points and 71 comments. A language model living in a boot sector reached 106 points, and WhatsApp MCP running free on Cloudflare reached 157.

### Key trend signals with data citations

The breadth of the cohort is the signal. An emergency assistant, a general assistant replacement, a finance app, an embedded toy and a messaging bridge are five different product categories sharing one technical stack, which is the shape of a platform shift.

Llmspeed's 331 points and 86 comments is the most commercially useful signal. It addresses a user who is deciding what hardware to buy in order to run local models — a user who has already concluded that local is worth investing in, and who represents the same audience one year before they ship a product.

Every one of those products rebuilt the same infrastructure. Downloading a multi-gigabyte weight file over an unreliable connection, resuming it, choosing the right quantisation for the detected hardware, and reapplying platform code-signing rules is weeks of work that appears nowhere in a product's differentiation.

### Why now

Expectations of software updates changed. Users are accustomed to incremental app updates measured in megabytes, and a model refresh that re-downloads three gigabytes reads as a bug rather than a release.

Commercial on-device products also arrived, which created the entitlement requirement. A model sold as part of an app needs offline enforcement, and there is no service that issues signed tokens for a locally running model.

### Market maturity

Early, and unusually welcomely so for an infrastructure product: the cohort is visible, the pain is uniform, and no incumbent has a reason to serve it. Hugging Face's business is research hosting; Ollama's is runtime; neither sells delivery with entitlements.

### What this means for a solo founder

Sell the measurement. Migrating three of October's on-device products and publishing their before and after numbers is worth more than any feature list, and it costs three conversations.

Survival LLM, an AI assistant and emergency library running entirely on local hardware, reached 543 points and 204 comments.

On-device AI becomes a consumer product

Hob, an open-source on-device alternative to Claude Cowork and ChatGPT, reached 179 points / 43 comments.

Local replacements for cloud assistants

Pulse Local, a Mac budgeting app whose models run on the machine, reached 122 points / 71 comments — local as a default rather than a feature.

Local models in ordinary apps

Llmspeed, which lets you feel local model speed before buying hardware, reached 331 points / 86 comments.

Local Model Delivery Service for Desktop Apps