ZYR3 routes to 34 specialists with 1M context and local execution

ZYR3 is a multi-AI model that functions as an auto-router selecting from 34 specialist models based on task type, context, and token budget. It supports up to 1 million tokens of context and includes 14 language specialists along with an ULTRA Coding Prompt.

• The system features a microVM sandbox for automatic code execution, testing, and self-refinement.
• Two editions are available: ZYR3 Lite offers unlimited usage, while ZYR3 Full provides a 5 million token budget.
• Built on the GLM-family base with an MIT license, it can be served via Transformers or vLLM.
• Benchmarks show performance of 6/7 on hard SWE Pro Terminal tasks and 5/5 on heavy-debug and codebase-exploration suites.

Users can run ZYR3 locally using standard Hugging Face pipelines or the vLLM server, with a provided script to bridge OpenAI-compatible clients.