Local LLM Demo — Docs
Reference

Models and endpoints

Base URL

https://llm-34a13a96.bunsenbrenner.org

Model

LiteLLM model name Backing model Where it runs
local-qwen3-coder qwen3-coder:30b (MoE, ~3.3B active params, Q4_K_M, ~18GB on disk) Ollama, this origin’s NVIDIA TITAN RTX (24GB VRAM)

Other model names configured on this same LiteLLM instance (cf-llama-70b, cf-llama-free, and the *:cloud fallback chain) are cloud-routed and not accessible to keys scoped to local-qwen3-coder — see Add an authorized person.

Endpoints

Path Shape Notes
POST /v1/chat/completions OpenAI Standard chat completions.
POST /v1/messages Anthropic LiteLLM’s Anthropic-format passthrough; what Claude Code talks to.
GET /health/liveliness No auth required; used by the tunnel’s own health checks.

measured Verified live 2026-08-29: /health/liveliness answers 200 ("I'm alive!") unauthenticated; /v1/chat/completions and /v1/messages both exist and return a 401 auth error when called with no key.

Auth

Authorization: Bearer sk-... (OpenAI-shaped calls) or x-api-key: sk-... + anthropic-version: 2023-06-01 (Anthropic-shaped calls). Same key works for both.

Known deployment-specific quirks

Found an error, or something that didn't work as documented? Open an issue →