Models & providers
Wednesday is not a model host. It routes to any endpoint that speaks the OpenAI chat API, and it keeps your keys on your machine.
Local models#
The local scan probes the usual servers. No key, no account.
| Server | Address |
|---|---|
| Ollama | 127.0.0.1:11434 |
| LM Studio | 127.0.0.1:1234 |
| llama.cpp server | 127.0.0.1:8080 |
| vLLM | 127.0.0.1:8000 |
| Any local server | 127.0.0.1:1337 |
| Wednesday local proxy | 127.0.0.1:18090 |
A vLLM server on a non-default port works too: paste the base URL. The same applies to a tunnel (ngrok, Cloudflare) that exposes an OpenAI-compatible endpoint.
Hosted providers#
Any OpenAI-compatible API works: OpenRouter, NVIDIA NIM, DeepSeek, Groq, Together, Mistral, or a self-hosted vLLM. Routes and keys live in ~/.wednesday/.env on your machine only — never in the repo, logs, receipts or the website.
| Backend | How |
|---|---|
| OpenRouter | Hosted catalog; OPENROUTER_API_KEY. |
| NVIDIA NIM | Hosted inference; NVIDIA_API_KEY. |
| Any OpenAI-compatible API | Base URL, key and model id. |
| Local proxy | Anything on 127.0.0.1:18090. |
Which model takes which seat#
A model pool comes from FACTORY_MODEL_POOL (id or id@baseURL, comma-separated), else the crew slots, else the single LOCAL_LLM_MODEL.
With team-up on (FACTORY_RELAY=1), the first team-up round probes every pooled model on five capabilities — reasoning, coding, vision, review, complete — and keeps the scores. Each seat then goes to the model that scores best for that capability. One strong model can hold the planner, the coder and the reviewer at once. A round’s wins and losses fold back into the scores, so the next round prefers the crew that proved itself. Routing is by measured capability fit.
FACTORY_MODEL_POOL=qwen/qwen3-coder,openrouter/anthropic/claude-sonnet-4
FACTORY_RELAY=1
Context window#
Two numbers matter, and the smaller one wins. The registry gives the model’s own ceiling — Qwen2.5-Coder-3B is a 32K model — while the server gives the deployment’s real window: the Colab T4 box that serves it runs a 13,312-token window. The smaller one wins.
A prompt that does not fit is trimmed in the middle, keeping the head and the tail, instead of being cut off at the end.
The operator’s roster#
These are the seats and local weights the reference roster runs. Serve them from any OpenAI-compatible server and point FACTORY_CREW_<SEAT>_BASE_URL at it.
| Seat | Model | GGUF weights |
|---|---|---|
UI/UX design (vision) | Qwen2.5-VL-3B | unsloth/Qwen2.5-VL-3B-Instruct-GGUF |
Architecture (architect) | Granite 3.1 2B | bartowski/granite-3.1-2b-instruct-GGUF |
Core coding (coder) | Qwen2.5-Coder-3B | Qwen/Qwen2.5-Coder-3B-Instruct-GGUF |
Autocomplete (autocomplete) | DeepSeek Coder 1.3B | TheBloke/deepseek-coder-1.3b-base-GGUF |
Bug hunting (bug-hunter) | Llama-3.2-3B-Instruct | unsloth/Llama-3.2-3B-Instruct-GGUF |
Security audit (security-audit) | Granite Guardian 3.1 2B | DevQuasar-11/ibm-granite.granite-guardian-3.1-2b-GGUF |
Quantisation is mandatory: download the Q4_K_M or Q5_K_M GGUF. You get most of the coding performance at half the hardware resource usage.
What a run costs#
Tokens, wall clock and dollars are recorded per seat. For a locally hosted model the real cost is GPU-minutes, so wall clock is on the same line. Wednesday will not invent a price: set your own contract price, or the dollar column stays honestly empty.
See the receipt for the pricing variables.