iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
The website operator usually pays for storing and delivering the model file, through its hosting or CDN provider. The visitor pays through their own connection, the storage space the file occupies, and the device time spent running it. How much the operator actually pays depends on the provider, the plan, cache behavior, traffic, and the regions visitors come from. There is no single cost per download that holds for every site.
The 50 MB figure is a scenario, not a standard size for browser AI tools. Model files, runtimes, and setup assets vary from one tool to the next, and the size of a browser-managed model can change as the browser updates it.
What the 50 MB figure does and does not describe
A DEV Community article that surfaced in search describes a download of about 44 MB of model plus a 5.95 MB runtime and setup package. The page could not be opened when this article was prepared, so that breakdown is the author’s stated example, not a verified measurement. Treat it as one illustration of how the total can be built up.
Other published figures describe different things entirely:
#1 Best Overall
| Figure | What it measures | Source and date | Limit on how it applies |
|---|---|---|---|
| About 50 MB total | The scenario in the article title | Title framing; no independent source | Not a general benchmark |
| About 44 MB model plus 5.95 MB runtime and setup | The author’s stated breakdown | DEV Community article; publication date not stated | Page could not be opened; not independently confirmed |
| 50 to 200 MB | Downloaded web bundle for bitHuman’s WebGPU avatar feature | bitHuman WebGPU documentation, updated 2026-10-04 | One vendor’s avatar bundle, not a typical language model |
| Approximately 20 GB minimum free disk space | Eligibility for Chrome on-device generative AI downloads | Google Chrome Help, accessed 2026 | An environment requirement, not the size of the model file |
| At least 22 GB free volume space | Environment for Chrome’s Prompt API built-in model | Chrome for Developers Prompt API documentation, current at time of writing | An environment requirement, not the size of Gemini Nano |
Chrome also says the exact size of Gemini Nano may vary as the browser updates it. For a custom tool, the only size that matters for billing is the total bytes the visitor’s browser actually fetches, which you can measure on your own site.
Two kinds of in-browser AI, with different bill payers
“In-browser AI” covers two arrangements that treat the model file differently. Which one a tool uses determines who carries the delivery cost.
A custom application that hosts its own model weights
Here the site operator uploads the model weights, or a bundle that contains them, to its own origin or a CDN. The visitor’s browser requests that file, receives the bytes, and stores them for later use. The operator’s hosting and delivery arrangement is what generates the bill for the transfer. The visitor’s connection and device absorb the rest of the cost.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
A browser-managed built-in model
In Chrome’s Prompt API, the browser downloads Gemini Nano separately the first time an origin uses the API. The Chrome for Developers documentation states two points that matter for cost:
- “The network requirement is only for the initial download of the model.”
- “No data is sent to Google or any third party when using the model.”
These statements describe Chrome’s documented built-in model use. They do not apply to every browser AI website. In this arrangement the site does not serve the model weights from its own delivery path, so the operator’s bandwidth bill for the model file does not arise in the same way. The visitor still bears the initial download, the storage, and the local compute, and Chrome’s built-in requirements recommend an unmetered connection.
The operator’s bill
In a custom application, the operator may pay for several items, and each depends on the architecture and the provider’s terms:
- Storage of the model file and any versioned copies
- Request charges for each fetch of the file
- Bandwidth or egress for bytes sent to visitors
- Hosting or CDN plan fees, including any included quota
- Engineering and operations time to configure caching and deployment
How major providers document delivery charges
The three providers below document different billing structures. None of them sets the rule for a site hosted elsewhere, and the rates depend on region and plan, so check each provider’s current price page before estimating.
Recommended Free Tools
| Provider and service | Documented billing element | What it means for a model download |
|---|---|---|
| Google Cloud, Cloud CDN | Bandwidth and HTTP/HTTPS request charges for CDN-served content | Each download to a visitor generates bandwidth charges and request charges; rates are not stated here |
| AWS, CloudFront | Viewer transfer charges under usage-based arrangements; flat-rate plans are also described | The bill depends on whether the account is on usage-based pricing or a flat-rate plan, and on the plan’s terms |
| Cloudflare, proxied domain and R2 | Bandwidth included in its proxied-domain example; free egress documented for R2 | Included bandwidth depends on the plan; R2 egress is documented as free, but storage and operation terms are separate and not covered here |
The visitor’s side
The visitor carries costs that never appear on the operator’s invoice:
- Data use. Chrome’s developer guidance states: “AI models can be large, which could lead to a large use of mobile data and device storage.” The statement is from Maud Nalpas, Kenji Baheux, and Alexandra Klepper, Chrome for Developers, published 2024-05-14.
- Wait time. The first inference cannot begin until the download completes.
- Local storage. The file stays on the device for later use, and Chrome’s on-device generative AI downloads carry the disk requirements listed above.
- Device compute and power. Running the model uses the visitor’s processor or graphics hardware. No quantified figure for the electricity this uses was established in the sources reviewed, so do not describe browser-side inference as costless.
Caching cuts repeat transfers, not the first one
Caching works at two levels, and they reduce different costs:
- Browser cache on the visitor’s device. bitHuman’s documentation states: “One download: the avatar’s web bundle (50–200 MB) downloads to the browser once, then comes from the cache.” This applies to that avatar feature. It avoids repeat downloads on the same device.
- CDN cache. A CDN may serve a stored copy, which reduces fetches back to the origin server. It does not stop the CDN from delivering the bytes to each viewer, and the CDN’s own billing rules still apply to that delivery.
Neither layer makes the first download free.
Estimating the operator’s exposure
The simplest way to size the cost is to separate the bytes the operator sends to visitors from the bytes it pulls from its origin. The first is incurred on every full download, and the second only on cache misses.
A worked example uses arithmetic only. If 10,000 new visitors each download a 50 MB model, that is 500,000 MB, or roughly 500 GB, of viewer delivery before any compression or caching of the origin. To turn that into money, multiply it by the viewer-transfer rate on your provider’s current price page for the regions your visitors are in, then add origin fetches and request charges.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Gather these inputs before estimating:
- The provider, the plan, and whether the account includes a bandwidth quota
- The regions visitors come from, because rates vary by geography
- The number of first-visit downloads per month
- The share of repeat visits served from the visitor’s browser cache
- The cache-hit ratio at the CDN and the number of cache misses that reach the origin
- The size of the file as served, after compression, plus any runtime or companion files
What a site operator can do
- Measure the total bytes the page fetches on first load, including runtime and companion files, and publish that number.
- Check the hosting or CDN provider’s pricing page for viewer transfer, request charges, and included quotas in the regions you serve.
- Use versioned file names and long-lived cache headers so that browsers and the CDN keep the file and do not refetch it on every visit.
- Tell visitors the download size and that the file will be stored on their device before the tool starts.
- Consider a browser-managed built-in model where it fits the task. The site then avoids serving the weights, but the browser and device requirements above apply, and the arrangement is specific to the browser and API that provide it.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
