OFFPEAK JOURNAL
Open models. Useful work.
Product · Product preview
OFFPEAK is being built for text workloads that can use an open model and adapt to available capacity. The first pilot is deliberately narrow: one approved model revision, one region, non-streaming text, and a clear accounting loop.
Start with the task
Extraction, classification, and summarization can be useful candidates. Define the expected output and build a representative evaluation set before choosing a model. An open model is suitable only if it meets your workload’s quality threshold.
3 concise points
source: supplied document
Know which model ran
The proposed service pins the weights revision, tokenizer, and chat template. Model substitution is outside the intended request contract. Licensing and the right to serve the chosen model must be established before paid access.
Use capacity deliberately
Available serving slots influence the quote, subject to a cost floor. Spare capacity can reduce the offered price, but no discount or latency outcome is guaranteed. Flexible queues can accommodate variation better than urgent interactive work.
Set a limit before execution
A quote binds the prompt, output ceiling, rates, and maximum charge. Wallet authorization carries the spending cap. Acceptance then reserves funds and capacity together, instead of letting dispatch get ahead of accounting.
Keep the result and record
A successful response is persisted before final charging. The unused reserve is released, and a signed receipt records the billed usage. Publication of a receipt batch follows separately from response delivery.
Start with the complete loop
The initial product must demonstrate deposit, capped call, receipt verification, and withdrawal. The current site is a preview of that product; it is not evidence that the paid service has launched.
Based on the supplied OFFPEAK version 1.0 product and build briefs. Read the overview ↗
← All articles