What is AI inference?

Every time you ask an AI assistant a question and it answers, you are running inference. Inference is the "use" phase of a machine-learning model: the model takes input it has never seen before, runs it through the fixed parameters it learned during training, and returns an output. No learning happens during inference. The model's weights do not change; it simply applies what it already knows to your specific prompt.

How is inference different from training?

Training and inference are two separate phases with very different costs and frequencies. Training builds the model: it processes enormous datasets, adjusts billions of internal parameters, and happens once (or occasionally, when a model is updated). Inference uses the finished model: it runs millions or billions of times, once per request, and never changes the parameters. A useful analogy is education versus work. Training is the years of study; inference is answering each question on the job afterward.

This split matters because the two costs land on different people. The labs that train frontier models absorb the gigantic one-time training cost. Anyone who builds a product on top of those models, including MiyoMind, pays the recurring inference cost for every message a user sends.

How does inference actually work?

For a large language model, inference happens token by token. Your prompt is converted into tokens (chunks of text), fed through the network, and the model predicts the next token, then the next, until the response is complete. Three things drive cost and speed:

• Prefill, where the model reads and processes your entire input prompt at once. Longer prompts and more context cost more here.
• Decode, where the model generates the output one token at a time. Longer answers cost more here, and this is what you watch when text streams in word by word.
• Context window, the total amount of text the model can hold in mind for a single call. Bigger context means more compute per request.

Two practical properties fall out of this: latency (how long you wait for the answer) and cost (how much compute the call burned). Latency depends on prompt length, output length, the size of the model, and how busy the serving hardware is. Cost scales with the number of input and output tokens, which is exactly why nearly every AI API prices by tokens processed.

Why is inference metered and charged?

Inference is metered because it is a real, recurring expense, not a fixed one. Every answer consumes GPU or accelerator time, and that compute is paid for whether you send one message or ten thousand. A subscription alone cannot cover unbounded usage, so AI products track how much each request actually costs and charge proportionally. This is the honest model: heavy users pay for heavy use, and light users do not subsidise them.

How does MiyoMind handle inference?

MiyoMind is a personal AI assistant you talk to inside WhatsApp, Telegram, Discord, and a web dashboard at miyomind.com. Behind every reply is an inference call to a frontier model. MiyoMind runs the open-source OpenClaw agent runtime, a model router called Hermes, and its own orchestration, memory, billing, safety, and routing code, then routes each request to the right model from OpenAI, Anthropic, Google, xAI, or Alibaba. Hermes picks the model based on the task, so a quick reminder does not pay for the most expensive model and a hard reasoning task gets the capacity it needs.

Because inference has a genuine cost, MiyoMind meters it with credits. One credit is worth roughly $0.005 of value, and credits track the actual model and tool usage of each request rather than charging a flat fee per message. The plans line up with how much inference you expect to run:

• Free, $0/mo, 100 credits every month, no card required, running on a shared direct-agent path with no dedicated container.
• Plus, $14.99/mo, 6,000 credits/mo, with one dedicated, sandboxed container.
• Pro, $39.99/mo, 18,000 credits/mo, with one dedicated, sandboxed container.
• Top-up packs (600/$3, 2,000/$10, 5,000/$25, 10,000/$50) when you need more inference without changing plans.

Paid users get their own isolated Docker container with no public internet egress, a read-only root filesystem, and zero stored external API keys, while integrations and memories are encrypted at rest with AES-256-GCM. The point of metering is fairness: you see your usage, the price reflects the compute each inference actually used, and there are no hidden surprises.