For current data times, read
data.window.end_date in the demand export and data.collection_finished_at in the supply export. Check the complete response schema and freshness before using a timestamp. Documentation does not represent a live data snapshot.What this dashboard can answer
It can compare observed OpenRouter demand and checked public model size. It can also compare current model supply and supply history. These measurements are separate.What this dashboard cannot answer
It cannot measure all market demand or prove that supply is short. It cannot select hardware, predict unit economics, or calculate the exact memory of a serving deployment. The short supply window does not measure long-term reliability.Observed OpenRouter demand
A model that is absent from a daily top-50 row has no separate detail. The dashboard does not convert that absence to zero demand. Visible demand is OpenRouter demand under these source limits. It is not total market demand.
Checked model-fact states
Missing values stay missing. They do not become zero. An older value does not replace a failed current check. The public reason states why the value is missing.
Weights and sequence state
Theoretical BF16 weights only
The value is the checked total parameter count multiplied by two bytes. It estimates complete-model weights. It does not show placement on each GPU.Smallest reviewed exact quantized checkpoint
This is the smallest stored size from the reviewed exact quantized checkpoints. Each checkpoint passed the bounded search and a human relationship review. Neither number is the complete loaded GPU memory. A stored checkpoint size is not the loaded quantized memory. Both numbers exclude engine, allocator, activation, temporary tensor, cache, concurrency, and workspace overhead.16-bit attention cache
The dashboard calculates theoretical attention-cache bytes for a batch size of one and a generic 16-bit data type. It uses the stated calculation sequence length. The record gives the basis, formula, included bytes, and excluded bytes. Nonstandard attention needs a reviewed architecture resolver. Recurrent or linear-attention state stays separate. A result is unavailable when required architecture fields are missing, unsupported, or in conflict. The record gives the exact reason. Weight quantization does not imply attention-cache quantization.Current supply snapshot
The current snapshot calculates provider and endpoint counts. It also calculates default prices, p50 latency, throughput, uptime, token limits, quantization, tool support, and implicit caching. It uses current OpenRouter rows. Null source values stay unavailable. It shows minimum, median, and maximum values for price, p50 latency, output throughput, and one-day uptime. Each calculation uses the current endpoint rows that report the value. Zero Data Retention coverage separates exact provider matches from unknown and ambiguous matches. Field coverage gives the number of current endpoint rows that report each field. A route quantization label is not a reviewed Hugging Face checkpoint. Conditional price overrides are counted but are not converted into the default-condition price.Public history fields
Price is a current snapshot because the public history record has no price field. Provider rows are private. The public output contains calculated model values only.