Prime Intellect says it launched Prime Inference, a serving product for frontier open-source models across multiple datacenters. In its launch posts, the company said the service was available that day with serverless endpoints and reserved capacity.
The immediate pitch is that Prime Intellect is opening up an inference stack it says it had already been using internally. In one getting-started post, the company said users can access GLM-5.3 through its API and “point any OpenAI SDK” at the service.
What Prime Intellect says is under the hood
In its technical blog post, Prime Intellect describes a stack that includes prefill/decode work, NVFP4 KV compression, and NIXL transfer work behind its GLM-5.3 deployment on GB200 NVL72. That makes the launch more than a bare product announcement: the company is also using it to show the infrastructure work it says powers the service.
Prime Intellect also says its production stack processes about 600 billion tokens daily on its own compute. In a longer post, Prime Intellect's Caoshuxun said internal inference usage is 600 billion to 700 billion tokens per day, after the company expanded use of its GLM-5.3 endpoint across agent research, RL rollouts, and synthetic data generation.
Those figures are central to the launch story because they are part of Prime Intellect's argument that Prime Inference is based on a system it already runs at significant scale. They are still company statements, not independent measurements in the supplied source packet.
Performance claims remain self-reported
Prime Intellect says its GLM-5.3 deployment ranks among the fastest GLM-5.3 endpoints on OpenRouter. In another launch post, it also claimed 100% uptime and a near-zero tool-call error rate.
The supplied evidence attributes those claims to Prime Intellect's own posts and blog, not to an outside benchmark, audit, or third-party test.
Early reaction showed interest, not verification
The supplied posts include positive reactions and amplification from employees and related accounts. One Prime Intellect employee wrote that the company had shifted many of its internal workload to its own fast GLM-5 endpoint and was now making it available to everybody. Prime Intellect also reposted a vLLM account message praising the launch and describing Prime Inference as GLM-5.3 served by vLLM at scale.
That response helps show interest around the project. It does not independently confirm Prime Intellect's throughput, uptime, or ranking claims.
For developers, the practical takeaway is simpler than the infrastructure jargon: Prime Intellect is offering Prime Inference as a live product with public API access and OpenAI-compatible usage, while the biggest performance claims around the service remain the company's own.
Comments