A product team sends a prompt to hosted weights and reads a completion back, without loading those weights onto its own machines. The provider accepts the call as an Application Programming Interface (API) request. That request travels as Hypertext Transfer Protocol (Secure) (HTTPS). The provider runs the forward pass on Graphics Processing Unit (GPU) hardware it operates.
Calling an open model such as DeepSeek usually works this way. The weights stay on the provider’s cluster. Downloading published weights and serving them yourself is a different job, because you then own capacity, drivers, and uptime. The pattern is most common for a Large Language Model (LLM), and the same hosted offer also covers embedding, speech, and image models in Artificial Intelligence (AI) products.
How Token Billing Works
The invoice follows the tokens a request moves. Fresh prompt text, a reused prefix, and generated text are three different lines. A token is a piece of text, smaller than a word, and hosts do not all split the same text the same way.
- Input tokens: The prompt the caller sends, including instructions, conversation history, tool definitions, and any passage fetched for the request. Text added by Retrieval-Augmented Generation (RAG) counts as input even though the documents were retrieved outside the model host.
- Cached input: A repeated prefix, such as a standing instruction or a long document, that the provider has already processed. That slice is billed as cached input, on its own line, rather than only as fresh input. Hosts do not share one cache rate, so one host’s cache price is not a market rule.
- Output tokens: The tokens the model generates in the reply. On a reasoning model, tokens produced before the visible answer are often billed as output too, so the output line can be longer than the text a user sees.
- Unit labels: Hosts quote a price per Million (M) tokens. The label 1M means one million tokens. A larger quote is priced per Billion (B) tokens. The label 1B means one billion tokens. Compare cards in the same unit.
What the Host Runs
The provider keeps the trained weights and performs the inference. The caller does not receive the weight file as part of the response. Read each host’s current price card for input, cached input, and output. A claim that one vendor is cheapest goes stale when that card changes, and it ignores context length and rate limits.
- Weights: The parameter file stays on the host for the life of the call. An open checkpoint can still be downloaded from the model maker and served elsewhere. The hosted call does not require that download.
- Hardware: The host runs the forward pass on GPUs it operates. Those machines are commonly NVIDIA accelerators. Replacing a failed card, or adding capacity for a busy hour, is the host’s work.
- Expert routing: Some open models use a Mixture of Experts (MoE) layout. The host stores the full weight set. One request may activate only a subset of the experts.
- Routing and direct hosts: A call can go to OpenRouter, which routes it to an underlying host, or straight to a host. Direct hosts include Together AI, DeepInfra, and Bitdeer AI Model Studio.
- Limits you do not set: The host enforces context length and rate limits. Some offers add a Service-Level Agreement (SLA) for uptime and support. That agreement is separate from the price per token.
Hosted Calls Compared with Rented Machines
Renting a GPU and loading the weights yourself is Infrastructure as a Service (IaaS). You pay for the machine while it is allocated, including hours with no traffic. A Platform as a Service (PaaS) host can run your container and still leave the model process to you. A finished product sold as Software as a Service (SaaS) hides the model behind the application. The hosted model sits between those offers: the weights are what you call, and the meter is tokens.
- No idle meter: A hosted call bills tokens that moved. A rented GPU bills the allocation, including quiet hours, whether or not a prompt arrived.
- No weight file in the response: The hosted path returns a completion. Serving a file you downloaded is the self-hosted path, with your own runtime, updates, and failures.
- The app still owns the prompt: Retrieval, instructions, and tool results are assembled before the call. The host does not become the product those steps belong to.
| Offer | What You Buy | What You Operate | How You Pay |
| Model as a Service | A completion from hosted weights | Prompts, retrieval, and product logic | Input, cached input, and output tokens |
| Infrastructure as a Service | GPU time on a machine you rent | The weights, the runtime, and scaling | Hours the machine stays allocated |
| Platform as a Service | A place to run your own software | The model process you deploy | Platform charges, not a token meter |
| Software as a Service | A finished application | Accounts and settings | A subscription or a seat fee |
| Self-hosted weights | No hosted model | Hardware, drivers, and the serving process | Equipment, power, and staff |
Additional Acronyms for MaaS
- MaaS - Marketing As A Service
