Open-weight models let you download and run the model yourself, under its licence. Closed models keep the weights private and provide access through a service or API. Start with a managed API when speed matters most. Consider self-hosting when data control, predictable demand or deployment restrictions justify operating it. Neither choice makes your brand appear in public AI answers.
I'm Brenden House, founder of Searchmaxxed. This is my decision guide for marketing and growth leaders, rather than a ranking of model vendors.
In summary
- Choose the deployment separately. Open weights can run through a hosted API, on rented cloud hardware or inside your own infrastructure.
- Compare completed work. Token rates leave out integration, review, idle hardware and the cost of fixing wrong answers.
- Read the exact licence. Downloadable weights don't transfer ownership of the developer's underlying intellectual property.
- Keep search visibility separate. Your private assistant and the public engines recommending suppliers read different sources.
What is the difference between open-weight and closed models?
Open-weight models expose their learned parameters for download; closed models do not. Those parameters, called weights, help determine the model's output for a given input.
Hosting is a separate decision. An API is the connection your software uses to request an answer, rather than a description of who owns the model.
| Deployment | Best starting point for | Main cost | Who operates inference? |
|---|---|---|---|
| Closed model through an API | Broad tasks without model infrastructure | Usage, integration and review | Service provider |
| Open-weight model through a hosted API | Testing open models without managing servers | Usage, integration and review | Hosting provider |
| Open-weight model you self-host | Deployment control and suitable repeat workloads | Hardware, operations and review | Your team or contracted operator |
Inference means running a trained model to produce an answer. Self-hosting changes who operates that process, including its security, capacity and updates.
The distinction matters commercially. A hosted open-weight model still sends your requests to another company, even though you could move those weights elsewhere under the licence.
Is open-weight the same as open source?
No. Open-weight describes access to weights; open source involves broader freedoms and the materials needed to modify the system.
The Open Source Initiative's definition requires data information, code and parameters under appropriate terms. Download access alone falls short.
Read the distinction carefully. The definition requires detailed training-data information; it does not require publishing every original training file, including material that cannot legally be shared.
Do you own an AI model when you download its weights?
Usually, you hold rights under a licence rather than ownership of the original model. Your rights depend on the particular release and its terms.
Read by version. A developer can release different models under different licences, while offering separate commercial terms for its hosted services and enterprise customers.
These examples illustrate the differences, rather than recommend particular models. Read each licence alongside the terms for your intended deployment.
| Model or service | Official wording | What to check before deployment |
|---|---|---|
| OpenAI gpt-oss | “These models are supported by the Apache 2.0 license”. | Licence notices, distribution obligations and the exact downloaded release |
| Meta Llama 3.1 | “non-exclusive, worldwide, non-transferable and royalty-free limited license” | Attribution, acceptable use and commercial conditions |
| Gemma models covered by Google's Gemma Terms | “Google claims no rights in Outputs you generate using Gemma”. | Covered versions, prohibited uses and redistribution requirements |
| OpenAI business services | “Customer: (a) retains all ownership rights in Input; and (b) owns all Output”. | Applicable-law qualification, output similarity and restrictions on use |
The last row concerns content. OpenAI's agreement qualifies those rights as between customer and provider, to the extent applicable law permits; it does not transfer the model weights.
Llama has specific conditions. Its licence requires “Built with Llama” attribution in specified distribution situations, and contains additional commercial terms for exceptionally large services.
Gemma also needs version checking. Google's terms page directs Gemma 4 users to Apache License 2.0, rather than applying the older Gemma terms across every release.
For procurement, separate the assets. Record rights to the base model, fine-tuned weights, input documents and output, then record who controls the application code and hosting account.
Ask legal to review the actual deployment. A licence summary cannot settle rights for a customer-facing product, a restricted industry or a model trained using another provider's outputs.
Which costs less: a closed API or a self-hosted LLM?
Neither is automatically cheaper. An API avoids dedicated model infrastructure; self-hosting can spread operating costs across steady demand, provided the model meets your required quality and speed.
Compare the whole job. A low inference bill loses its appeal when editors spend longer correcting claims or customers wait too long for a usable answer.
A token is a unit of text a model processes. Token billing therefore measures model usage, not whether the resulting work helped your team.
Use this calculation:
Cost per accepted task = total delivery cost ÷ tasks that meet your acceptance criteria.
| Cost to include | Managed API | Self-hosted open-weight model |
|---|---|---|
| Model execution | Input, output, caching and any tool charges | Hardware purchase or rental, power and capacity |
| Quiet periods | Usage-dependent, subject to your contract | Idle capacity still needs funding |
| Integration | Application, retrieval, access controls and testing | The same application work, plus serving infrastructure |
| Operations | Provider management, monitoring and change testing | Patching, scaling, backups, incidents and model updates |
| Quality | Review, retries and corrections | Review, retries and corrections |
| Exit | Rewrite provider-specific integrations | Move weights, runtime, data and dependent services |
Set a common quality threshold first. Then measure the cost of meeting it across representative work, including complicated requests, failed responses and the busiest periods you need to support.
Keep currencies consistent. For Australian teams buying US cloud services, record exchange-rate assumptions and tax treatment instead of comparing an A$ budget with a US$ headline rate.
When does self-hosting start making sense?
Self-hosting deserves a test when you have predictable work and someone who can operate the deployment. Control requirements can justify it even without savings.
Start with a bounded task. Classifying product questions against an approved taxonomy is easier to evaluate than replacing every content, analytics and support tool across the marketing department.
An API remains a useful baseline. Test a hosted open-weight version before committing to hardware, then repeat the evaluation on the actual configuration you intend to run yourself.
Does self-hosting keep customer data private?
Self-hosting can keep inference inside infrastructure you control, but the complete application determines where data goes. Check every external service, not just the model.
Follow a real request. It may touch an embedding service, document store, search tool, monitoring platform and support log before the customer receives an answer from your application.
A private GPU changes none of those connections by itself.
Closed APIs also differ. OpenAI's API data guidance says it does not train on API data unless the customer explicitly opts in.
Training and retention are separate. The same guidance describes default abuse-monitoring retention of up to 30 days, with longer retention possible under its stated exceptions.
Zero Data Retention needs approval. Eligibility and endpoint behaviour matter, including features that still retain application state, so check the configuration rather than treating the label as universal.
For enterprise or public-sector procurement, establish:
- where inference, storage and backups run;
- who can access prompts, retrieved documents and logs;
- which data the provider retains, and for how long;
- whether external tools or support access cross the approved boundary;
- how deletion, incidents and customer access controls work.
Ask your security team first. An Australian or US hosting region alone cannot establish compliance with your organisation's contracts, sector rules and obligations for the data involved.
Which model should marketing teams use for content, analytics and customer-facing tools?
Choose against the task, using the same source material and acceptance criteria for each candidate. Model family alone cannot establish fitness for your work.
Make failure visible. A polished paragraph can still invent a product capability, while a fluent analytics summary can reverse the meaning of a conversion-rate change in your report.
| Work | What to test | Sensible deployment starting point |
|---|---|---|
| Public content drafts | Product accuracy, source support, voice and editing time | Managed API, then compare open-weight candidates |
| Search and campaign analytics | Correct calculations, data permissions and repeatable conclusions | Approved environment with read-only access to required data |
| Customer-facing product assistant | Correct citations, access boundaries, refusal and handover | Retrieval-backed application with monitored responses |
| Confidential manufacturing documents | Permission checks, terminology and permitted data movement | Approved private deployment or compliant managed service |
For analytics, let tested code calculate. Ask the model to explain verified results, with the underlying tables available, rather than trusting a plausible sentence to establish that the arithmetic is right.
For customer-facing tools, test missing answers. A system that asks for help when the documentation is silent can serve customers better than one that confidently invents a feature or commitment.
Hybrid deployment is another option. Keep suitable repeat tasks on an open-weight model and route harder requests to a managed service, provided your data rules permit that transfer.
Do you need fine-tuning or retrieval?
Fine-tuning changes a model's parameters through additional training. Retrieval-augmented generation, or RAG, supplies relevant source material when the application answers a question.
They solve different problems. For current product specifications, start by testing retrieval against approved documents before deciding that the model itself needs further training to handle the task.
AWS's RAG explanation describes adding an organisation's knowledge without retraining the model. Its guidance also covers restricting retrieval according to user permissions.
Keep the source trail. If an assistant answers from an outdated product sheet, changing model vendors may leave the actual problem untouched because the retrieved evidence remains wrong.
Will owning your AI model improve your brand's AI search visibility?
Owning or self-hosting a model does not update the models or search systems your customers use. Your internal assistant and public AI recommendations are separate problems.
Start with the public evidence. Searchmaxxed's AI search work improves the pages, technical access and supporting sources that help engines understand your company and answer relevant buyer searches.
Google makes the distinction concrete. Its AI search guidance says there are “no additional requirements to appear in AI Overviews or AI Mode”.
Pages still need eligibility. Google requires indexing and snippet eligibility for supporting links, alongside familiar SEO work such as crawl access, internal links and visible content matching structured data.
OpenAI separates search from training. Its crawler documentation lets site owners allow OAI-SearchBot for search while disallowing GPTBot for foundation-model training.
That matters for your policy. You can distinguish public search access from training permission without pretending either setting gives you control over which brands an engine recommends to customers.
Our own evidence shows the gap between retrieval and citation. This comes from our Australian agency-buyer panel on 2 October 2026:
We put 300 Australian agency-buyer questions to ChatGPT three times each, with search on. It pulled 17,870 search results into those 900 answers and kept 4,389 as sources. Being found is the easy part. About three in four results it reads never make the answer.
That observation concerns the panel. It does not establish a universal citation rate, or show that buying one model family improves another engine's recommendations for your brand or products.
The practical distinction is simple. Self-hosting can change how your organisation answers; LLM SEO concerns whether other people's AI tools can find and recommend your organisation.
What would this look like for a B2B SaaS company?
Consider a hypothetical SaaS company expanding into Australia and the US. Its product assistant needs current integration documentation, while prospects need to discover the product through search.
Separate the workstreams. The engineering team might test a private retrieval-backed assistant, but that does not answer a prospect asking ChatGPT which platforms support their existing software.
For the public question, Searchmaxxed builds integration and comparison pages around verified product facts. We connect those pages through internal links and fix crawl and indexing problems.
Then we track the result. Do the relevant search and AI answers name the product, cite its pages and bring visitors who make enquiries or start meaningful sales conversations?
Same company, different work. A successful internal assistant cannot substitute for public pages that explain the product clearly enough for unfamiliar customers and external search engines to evaluate it.
How should you make the decision without getting locked in?
Keep your source documents, test cases and application interfaces portable from the start. Access to weights helps, but provider-specific tools and infrastructure can still make switching expensive.
Write a short decision record:
- Define the job and its acceptable answer.
- Name the data the application may access.
- Confirm the exact model licence and service terms.
- Compare candidates on representative tasks and difficult failures.
- Measure cost per accepted task, including review and operations.
- Test a replacement before signing off the architecture.
Keep the visibility budget separate. Our technical SEO work addresses crawl access, indexing and website structure, none of which an internal model migration fixes by itself.
Is Searchmaxxed right for your team?
Searchmaxxed fits teams that want customers to find them through Google and AI answers. Our role here is SEO and AI search optimisation, rather than selling model access.
We build and improve the public website. That includes product and industry pages, comparisons, schema, internal links and the technical fixes that help search engines reach the content.
Our AI search optimisation service also checks how ChatGPT, Gemini, Perplexity and Google's AI answers describe your business. We connect that evidence to work on the site and supporting sources.
Choose scope around the market. Growth System combines page releases with technical SEO and AI visibility tracking; Agentic Search Engineering adds broader page families, digital PR and qualified link campaigns.
Custom & Enterprise can cover multiple markets, dedicated engineering, custom reporting and original first-party research. You own the website, repository, content and data from the start.
| Searchmaxxed is a good fit if | Look elsewhere if |
|---|---|
| You want customers to find your company through Google and AI answers | You need GPU operations or model-licensing advice |
| You want public pages and technical SEO changes implemented | You want strategy advice with no changes to your site |
| You need SEO and AI search work across multiple markets | You only need a private internal assistant |
Frequently asked questions
Why are open-weight models cheaper?
They can reduce dependence on a single inference supplier and support competitive hosting options. Self-hosting still carries hardware, engineering and operating costs, so compare accepted work rather than token rates alone.
What is an example of an open-weight model?
OpenAI's gpt-oss is an example. Its official page offers downloadable models under Apache 2.0, with guidance for local deployment and fine-tuning.
How far behind are open-weight models?
There is no fixed gap. Compare specific versions on your actual tasks, because a general benchmark cannot establish whether either model handles your product facts, documents or customers well.
Are Chinese models open-source or open-weight?
The model release decides. A developer's country does not establish whether it publishes weights, training code or data information, or which licence governs commercial use and distribution.
Can you self-host a closed model?
Usually not without a specific vendor arrangement. Public API access does not include downloadable weights; confirm any private-deployment offer in the provider's actual service terms.
Does fine-tuning make ChatGPT recommend your company?
Fine-tuning your own model does not change public ChatGPT. Build accurate, accessible public sources and assess the answers customers see, rather than treating private training as public search optimisation.
Is a hosted open-weight API the same as self-hosting?
No. A hosting provider operates the model and receives your requests; self-hosting puts that operation under your control, subject to the infrastructure and services you choose.
If the missing result is customers finding your company through search, start with the public website and the questions those customers ask.