ASP DEV

Project

Pi.dev and local-model orchestration on an RTX 5070 Ti

A study of local Pi.dev orchestration: model roles, OpenAI-compatible providers, and a reproducible RTX 5070 Ti setup.

Pi Coding AgentOllamaLM StudiovLLMOpenAI-compatible APIRTX 5070 Ti
Demo

The constraint

One RTX 5070 Ti sets the boundary for this experiment. I am testing how Pi.dev can split work across local models, rather than asking one model to carry the whole workflow.

What I am testing

  • Ollama, LM Studio, or vLLM through models.json and OpenAI-compatible APIs.
  • Separate roles: a lighter model for navigation and classification, a stronger one for code, and a review pass.
  • RPC/JSON and SDK modes, task queues, context boundaries, and tool permissions.
  • Latency, VRAM pressure, context length, output quality, and fallback behavior when an endpoint is unavailable.

The result I want

A setup someone else can reproduce: which tasks fit on a 5070 Ti, when routing helps, and how to switch to another local endpoint. Benchmarks and exact model choices will come after the measurements.