Inference for open-weight models,
at interactive speed. beta

Register sevenk.ai as a custom provider in ~/.config/opencode/opencode.json, then authenticate once via /connect (choose sevenk.ai → key) or by setting SEVENK_API_KEY on the OpenCode server process. Models are discovered from the endpoint automatically.

{
  "$schema": "https://opencode.ai/config.json",
  "model": "sevenk.ai/qwen3.8-27b",
  "providers": {
    "sevenk.ai": {
      "name": "sevenk.ai",
      "package": "@opencode/ai/providers/openai-compatible",
      "settings": {
        "baseURL": "https://sevenk.ai/v1",
        "apiKey": "{env:SEVENK_API_KEY}"
      }
    }
  }
}

pi does no server-side discovery, so list the models in ~/.pi/agent/models.json. contextWindow comes from GET /v1/models; maxTokens is a client-side cap you choose.

// ~/.pi/agent/models.json
{
  "providers": {
    "sevenk": {
      "baseUrl": "https://sevenk.ai/v1",
      "api": "openai-completions",
      "apiKey": "sk-gw-…",
      "compat": { "supportsDeveloperRole": false },
      "models": [
        { "id": "qwen3.8-27b", "reasoning": true, "input": ["text","image"], "contextWindow": 262144, "maxTokens": 32768 },
        { "id": "qwen3.8-flash", "reasoning": true, "input": ["text","image"], "contextWindow": 179200, "maxTokens": 32768 }
      ]
    }
  }
}

In Cline, Roo Code or Kilo Code pick the OpenAI Compatible provider (not the built-in Ollama or LM Studio ones) so inference rides /v1/chat/completions and the model list comes from /v1/models.

Provider      OpenAI Compatible
Base URL      https://sevenk.ai/v1
API key       sk-gw-…
Model         qwen3.8-27b

Enter the model id by hand if it does not appear in the picker.

1. Settings → Models → Override OpenAI Base URL = https://sevenk.ai/v1
2. Add your API key, then Add Custom Model for each model id and press Verify
3. Select the model for Cursor Agent (or chat) in the model picker

Do not prefix model ids with claude-*: Cursor rewrites those to Anthropic's wire format.

Codex CLI does no discovery, so pin the model and point the provider at the gateway in ~/.codex/config.toml, then export SEVENK_API_KEY.

# ~/.codex/config.toml
model = "qwen3.8-27b"
model_provider = "sevenk"

[model_providers.sevenk]
name = "sevenk.ai"
base_url = "https://sevenk.ai/v1"
wire_api = "responses"
env_key = "SEVENK_API_KEY"

Anything that speaks OpenAI works: SDKs, Aider, LobeChat, n8n, scripts. Stream for long generations; errors keep the OpenAI shape.

curl https://sevenk.ai/v1/chat/completions \
  -H "Authorization: Bearer sk-gw-…" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-27b",
    "stream": true,
    "messages": [{ "role": "user", "content": "hello" }]
  }'

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://sevenk.ai/v1",
  apiKey: process.env.SEVENK_API_KEY,
});

const stream = await client.chat.completions.create({
  model: "qwen3.8-27b",
  stream: true,
  messages: [{ role: "user", content: "hello" }],
});