Guides
- Quickstart
- Authentication and keys
- Key protection
- Models, aliases and fallbacks
- Automatic model choice (silicon/auto)
- Chat completions
- Streaming
- Structured outputs and tools
- Logprobs
- Vision
- Responses, Claude and Ollama
- Embeddings, rerank, audio, images
- Files and batches
- Documents and search
- LoRA adapters
- Project tags and costs
- Prefix caching
- Headers, rate limits and spend caps
- Errors
- Privacy and retention
- Verifiable receipts
- Data residency
- Pricing
- Billing
- Pay per request (x402)
- On-chain oracle
- Agent credit (backed compute lines)
- Status
- SDKs
- Swift SDK
- Integrations
- Run a node
- Linux agent
- Compute API
- Architecture
- Changelog
- Glossary
API reference
Guide
Glossary
The words this site and the docs use, in plain terms. Where a term only applies to a workload that is not available yet, the definition says so.
- API key
- The secret your code sends with each request. It names your organization, carries limits and is what usage is billed to. Silicon Network keys start with sn_.
- Balance
- The prepaid credit on your account, in US dollars. Usage is charged to it; you top it up with stablecoins or a card.
- Decentralized network
- A network of machines run by many independent parties rather than one company. Planned as a source; none is connected yet.
- Dedicated server
- A whole physical machine rented to one customer, with nothing else running on it.
- Fine-tuning
- Adjusting an existing model on your own examples instead of training one from scratch. Often done with a small adapter (LoRA) that is stored and loaded next to the base model.
- Gateway
- The public API at api.siliconnetwork.io. It checks keys and balances, picks a node, streams the answer back and counts usage.
- GPU cloud
- A company that rents out GPU machines, often by the hour, sometimes as a marketplace of other people's hardware. Planned as a source; none is connected yet.
- GPU hour
- One GPU used for one hour: the usual unit for pricing training and rendering.
- Hash rate
- How many hashes a machine computes per second. It measures mining speed, not whether mining pays.
- Inference
- Running a trained model on new input to get an answer. Every chat message, embedding or generated image is an inference request. The one workload available today.
- Latency
- How long something takes to start answering. For a model, usually the time to the first token; for a server, the network round trip to it.
- LoRA adapter
- A small set of extra weights that changes how a base model behaves. Silicon Network can serve a registered adapter on top of its base model through the inference API.
- Metering
- Counting what each request used, such as input and output tokens, so it can be billed. On Silicon Network the gateway counts and the control plane charges.
- Node
- One machine in the network: an operator's Mac running the Silicon Network engine. No node has joined yet.
- On-demand
- Capacity you rent at a fixed rate for as long as you need it, which the provider does not take back while you pay for it.
- Operator
- A person or company that joins their own machines to Silicon Network and runs the engine on them. Today operators run Macs; payouts to operators are off.
- Proof of work
- A way for a blockchain to agree on its next block: machines compute hashes until one finds a valid one. Mining is that computation. Many cloud providers restrict it in their terms.
- Public cloud
- One of the large general-purpose clouds, with servers, storage and networking in many regions. Planned as a source; none is connected yet.
- Quantization
- Storing a model's weights with fewer bits, such as 4 or 8 instead of 16. The model needs less memory and runs faster, at some cost in quality.
- Receipt
- A signed statement of what served a request: the node signs what it produced and the gateway countersigns how it was routed. Optional, per request.
- Region and data residency
- A region is where a machine physically is. Data residency holds requests to chosen regions or countries, and refuses a request rather than serve it elsewhere.
- Relay
- The service that lets a node behind a home or office router serve requests. The node dials out to the relay, so nothing has to be opened inbound.
- RPC node
- A server that answers requests about a blockchain and passes transactions to it. Bots that trade on a chain talk to one constantly, so the distance to it matters.
- Spot or interruptible capacity
- Spare machines a provider rents out for less, on the condition that it can take them back at short notice. Suits jobs that save checkpoints and can resume; not suited to serving live traffic.
- Throughput
- How much work is done per unit of time: tokens per second for a model, frames per hour for a render job. High throughput and low latency are different goals and can pull against each other.
- Token
- The unit a language model reads and writes: a word or a piece of one. Inference is billed per input token and per output token.
- Training
- Building a model's weights from data, usually on many GPUs for hours or days. Not available on Silicon Network yet.
- VPS
- Virtual private server: a virtual machine with its own operating system, address and fixed share of a host's CPU, memory and disk.
- VRAM and unified memory
- VRAM is a GPU's own memory; a model must fit in it to run fast. Apple Silicon has unified memory instead, one pool shared by the CPU and the GPU, so a Mac with a lot of memory can hold a large model.
- Warm model
- A model already loaded in a node's memory, so it can answer without first being read from disk.
Silicon Network›Docs›Guides›Glossary50 items