isoquant

Isoquant is an inference platform that runs open models at low cost and latency behind an API intended to slot into an existing stack. The vendor describes engineering across four layers: architecture aware model optimization including mixed precision, GPU kernel and memory engineering, accelerated decoding including speculative decoding, and workload aware serving with KV cache and load balancing. It states that it is backed by Y Combinator. Developers use it by adding an Isoquant API key and changing the base URL, with the published example calling a chat completions endpoint at api.isoquant.ai. The featured model is GLM 5.3 Flash, for which the vendor publishes a P50 time to first token of 452 milliseconds, P50 end to end latency of 2.3 seconds and throughput of 158.9 tokens per second, measured against its own AgentX benchmark and OpenRouter data dated September 21, 2026. A second offering for companies is self optimizing inference. Isoquant connects to a customer's workload and success criteria, evaluates models, configurations and trained specialists, and deploys the winners. Published case studies describe a multi step order resolution agent whose task success rose from 15.8 percent to 98.7 percent through model selection and execution instructions. Published pricing for GLM 5.3 Flash is 0.027 dollars per million input tokens, 0.018 dollars per million cached input tokens and 0.49 dollars per million output tokens, with automatic prompt caching and no top up fees. The vendor does not publish a pricing page, and company engagements are arranged through a contact form on the site.