Chat

Two sources for LLM chat: ollama (Ollama-compatible) and vllm (OpenAI-compatible).

---

Ollama Source

Ollama-compatible API format.

POST /v2/ollama/api/chat

Request

{
  "model": "gpt-oss:20b",
  "messages": [
    {"role": "user", "content": "Hello"}
  ],
  "mode": "auto",
  "stream": false
}

Parameter	Type	Description
model	string	Model name (required)
messages	array	Conversation messages (required)
mode	string	auto, direct, or opengpu (default: auto)
stream	bool	Enable streaming (default: false)
think	string	Reasoning depth: low, medium, high
temperature	float	Randomness 0.0-2.0

Response

{
  "model": "gpt-oss:20b",
  "message": {
    "role": "assistant",
    "content": "Hello! How can I help?"
  },
  "done": true,
  "mode": "direct",
  "task_address": "direct:abc12345-...",
  "prompt_eval_count": 15,
  "eval_count": 8
}

Field	Description
mode	Actual mode used: direct or opengpu
task_address	Unique identifier. For opengpu mode, verifiable on ogpuscan.io

Streaming

curl -X POST https://relay.opengpu.network/v2/ollama/api/chat \
  -H "X-API-Key: YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-oss:20b", "messages": [...], "stream": true}'

Returns Server-Sent Events:

{"message":{"content":"Hello"}}
{"message":{"content":" world"}}
{"done":true}

Models (ollama source)

Model	Modes	Tiers
gpt-oss:20b	opengpu	all
gpt-oss:120b	direct, opengpu	pro/max
llama3.2:3b	opengpu	all
deepseek-r1:8b	opengpu	all

---

vLLM Source

OpenAI-compatible API format. Direct mode only.

POST /v2/vllm/v1/chat/completions

Request

{
  "model": "openai/gpt-oss-120b",
  "messages": [
    {"role": "user", "content": "Hello"}
  ]
}

Response

{
  "id": "chatcmpl-123",
  "object": "chat.completion",
  "choices": [{
    "index": 0,
    "message": {
      "role": "assistant",
      "content": "Hello! How can I help?"
    },
    "finish_reason": "stop"
  }],
  "usage": {
    "prompt_tokens": 10,
    "completion_tokens": 8,
    "total_tokens": 18
  },
  "mode": "direct",
  "task_address": "direct:abc12345-..."
}

Field	Description
mode	Actual mode used (always direct for vLLM)
task_address	Unique identifier for this request

Models (vllm source)

Model	Modes	Tiers
openai/gpt-oss-120b	direct	pro/max