---
title: THOX ZeroGPU - Coder
emoji: 💻
colorFrom: green
colorTo: gray
sdk: gradio
sdk_version: 5.49.1
python_version: "3.12"
app_file: app.py
pinned: false
short_description: Code generation and review on ZeroGPU
---

# THOX ZeroGPU - Coder

Serves **`thox-coder-7b`** behind an OpenAI-compatible endpoint on ZeroGPU, so that
Open WebUI can consume it as an ordinary "OpenAI API" connection.

| | |
|---|---|
| Base model | `Qwen/Qwen2.5-Coder-7B-Instruct` |
| LoRA adapter | `none (base weights)` |
| Served model name | `thox-coder-7b` |
| Hardware | ZeroGPU (`zero-a10g`), size `large` |



## Endpoints

- `GET  /v1/models` — model list
- `POST /v1/chat/completions` — chat completion, streaming and non-streaming
- `GET  /health` — liveness plus the resolved model identity
- `GET  /` — Gradio chat UI

## Connecting from Open WebUI

- **Base URL** — `https://tommytracx-thox-zero-coder.hf.space/v1`
- **API key** — the `THOX_API_KEY` value

## Authentication

`/v1/*` requires an API key, sent the way any OpenAI client already sends one:

```
Authorization: Bearer <key>          # X-API-Key: <key> also accepted
```

The key is the Space secret `THOX_API_KEY`. `/v1/models` is gated too — an open
`/v1/models` is how a bogus key gets mistaken for a working one. `GET /health` stays open
for liveness and reports `auth: required`. If the secret is unset the endpoint fails
**closed** (503), never open.

## OpenAI compatibility

Verified with the official `openai` Python SDK (see `conformance.py` in
`thox-zerogpu-spaces`), not by inspection:

| | |
|---|---|
| `client.models.list()` | ✅ |
| `client.chat.completions.create()` | ✅ `id` · `object` · `created` · `model` · `choices[].message` · `finish_reason` · `usage` |
| `stream=True` | ✅ `chat.completion.chunk` frames, `role` in the first delta, terminal `finish_reason`, `[DONE]` |
| multi-turn + `system` role | ✅ |
| `tools=[...]` function calling | ✅ OpenAI-shaped `tool_calls`, `finish_reason: tool_calls` |
| bad key | ✅ `401` → `openai.AuthenticationError` |
| malformed body | ✅ `400 {"error": {...}}` — not FastAPI's `422 {"detail": [...]}` |
| out-of-range `max_tokens` / `temperature` / `top_p` | ✅ clamped, not rejected |

Sampling ceilings are enforced by clamping rather than Pydantic bounds, because
`Field(le=...)` makes FastAPI answer `422 {"detail": [...]}` while the OpenAI API answers
`400 {"error": {...}}` — and a client that reads `error.message` sees nothing at all.

## Source

Generated from the shared template in `thox-zerogpu-spaces`. `app.py` is byte-identical
across every THOX ZeroGPU model Space — only `thox_space_config.py` differs.

Your AI. Your Data. Your Rules.&trade;