Frontier models, answered before you notice you asked.
Blink is an inference company. We take the largest models people actually want to run and serve them at latencies low enough that the waiting disappears — no spinner, no streaming used as an apology for a slow first token.
Nothing gets you under the threshold of perception on its own. Blink is built as one pass from the request's first hop to the last token, with every layer written against the same budget.
Draft models, tree-structured verification, and continuous batching are scheduled as a single decision rather than three independent ones, so speculative depth adapts to how loaded the accelerator actually is.
Attention and mixture-of-experts kernels are written per accelerator generation, not compiled down to a portable average. Weights stay resident; the KV cache never leaves high-bandwidth memory during a turn.
Serving racks sit inside the metros the traffic comes from. Speed of light through fiber is the one term in the budget we cannot optimize, so we shorten the distance instead.
A round trip on a 70B-class model, broken down. The interesting part is not the total — it is that no single stage is allowed to dominate.
Median end-to-end, 512-token prompt, 64 tokens out, same-metro client. The last row is the reflex the company is named after.
We are onboarding teams whose products break when inference is slow — voice, agents, interactive tools, anything with a person waiting at the other end.
hello@blink.infinity.inc