demo-chat

Local-first test bench for Ollama models on a shared-GPU Windows PC (i3-12100, Intel UHD 730 via Vulkan, 32 GB RAM): chat with any local model — including the ones OpenCode can't use as agents — with per-turn performance logging, history, model comparison charts, and a live VRAM view.

Quick start

cd app
bun src/server.ts          # http://localhost:3000

Requires Ollama at http://127.0.0.1:11434 with OLLAMA_IGPU_ENABLE=1 (see docs/Ollama-iGPU-Setup.md), otherwise everything falls back to slow CPU inference. Bun 1.4+ is the only dependency.

What's inside

Path	Contents
app/	the chat app (Bun + SQLite, zero npm deps) — see app/README.md
app/docs/	app docs: overview, usage, API, database, test reports
docs/	machine setup notes: Ollama on shared GPU, Farsi model tests, 9Router wiring + raw fa-tests/ outputs

Key findings so far

• Ollama ignores the Intel iGPU unless OLLAMA_IGPU_ENABLE=1 — then all models run at 100% shared GPU (docs/Ollama-iGPU-Setup.md).
• 16 models tested for Persian; 4 winners + 1 tuned farsi-gemma (docs/Farsi-Models.md). Gemma models lack OpenCode tools support, hence this app; aya-expanse:8b / command-r7b work as OpenCode agents.
• 9Router fronts Ollama as local/* for any OpenAI client (docs/9Router.md) — don't use prefix ollama, it collides with a built-in provider id.