By Jordan Vale
Training an open AI model is only half the job. Someone still has to run it for real users. On October 2, Prime Intellect said it opened that door with Prime Inference: pay-as-you-go endpoints for spiky demand, and reserved capacity for steady work.
A serving layer, not a new model. Prime describes it as the run-time piece of its open training stack. It speaks the same interface as OpenAI's tools, so existing software can point at a Prime address and key. Models run on Nvidia's current Blackwell hardware in more than one data center. The next Nvidia wave, Vera Rubin, is listed as coming later. Automatic failover is meant to move traffic when a site fails.
The first public test. The first model Prime put on the service is called GLM-5.3. It went live on OpenRouter on September 22. Prime says that endpoint ranks among the fastest GLM-5.3 options there, with a near-zero tool-call error rate and 100% uptime since launch.
Ink, not pencil: those uptime and error-rate figures are the company's claims, not an independent audit. Treat them as a vendor report until someone else measures the same traffic.
Prime says the platform already ran heavy internal work before the public launch: reinforcement-learning runs, synthetic data, evaluations, and long coding agents, at nearly a trillion tokens a day. It also says it has served large customer jobs since January. The stack it names combines Nvidia Dynamo, vLLM, Mooncake, and FlashInfer, built with partners including Inferact and Nvidia, with fixes sent upstream.
Still on the list. Cheaper batch and async jobs for big offline runs. Dedicated and one-click deployments on reserved capacity, including fine-tuned models from Prime's own training.
Why regular people should care
Open models only help if you can run them without buying a room of chips. The Monday stake for builders and small teams is practical: can you rent reliable inference the way you rent storage, turn it on when you need it, and reserve it when you do not want surprises?
What's next
Watch published pricing, outside speed and reliability checks, and whether the reserved and one-click options show up as promised. For now, the useful fact is narrow. Frontier open models have a new on-demand and reserved place to run, and GLM-5.3 is the first public example on OpenRouter. The reliability numbers are still Prime's.