PrimaLabs

The stack watches incoming requests, tries candidate changes, and ships one only when it measures faster with identical outputs. The site describes this as working without changing the model or the product and without keeping the data it learns from. The home page chart of speed rising with volume is labeled as conceptual, not measured data. PrimaLabs pitches this at coding agents that send many near-identical requests, multi-step agent pipelines that reuse long contexts, and products where model latency shapes the user experience.

The DeepSeek V4 Flash page lists $0.016 per million input tokens and $0.266 per million output tokens at a 90% cache hit rate, and compares that with other providers' effective prices on OpenRouter from July 2026. Cached input is billed at a fraction of list price, so the rate a customer actually pays depends on their hit rate. The API is OpenAI-compatible, and deployment options include US data residency, a customer VPC or on-premises hardware. There's a dashboard to try it, and larger customers can request a written benchmark of their own workload.

The privacy policy, effective September 17, 2026, says PrimaLabs doesn't use request inputs or outputs to train or improve models and doesn't keep that content after a request or session ends, with exceptions set out in the policy. The founding team comes from US national laboratory supercomputing work, and the company cites a first-place Artificial Analysis result for MiMo-V2.6-Pro. Those rankings change often, so check the current figures before relying on them.