opencode.md

OpenCode + Candle-vLLM

This guide connects OpenCode directly to candle-vllm through the built-in OpenAI-compatible /v1/chat/completions endpoint.

OpenCode -> Candle-vLLM (OpenAI-compatible)

1) Start candle-vLLM

cargo run --release --features cuda,nccl,graph,flashinfer,cutlass -- \
  --m Qwen/Qwen3.5-27B-FP8 \
  --d 0 \
  --prefix-cache \
  --p 8000 \
  --gpu-memory-fraction 0.7 \
  --enforce-parser qwen_coder

If you prefer FlashAttention, replace flashinfer with flashattn.

2) Discover the served model name

curl http://localhost:8000/v1/models

Use the returned id in the OpenCode config.

3) Configure OpenCode

Install OpenCode:

curl -fsSL https://opencode.ai/install | bash

Or:

npm i -g opencode-ai

Create ~/.config/opencode/config.json:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "local-candle-vllm": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Candle-vLLM Local",
      "options": {
        "baseURL": "http://localhost:8000/v1"
      },
      "models": {
        "qwen3-coder": {
          "name": "Qwen/Qwen3.5-27B-FP8"
        }
      }
    }
  },
  "model": "local-candle-vllm/qwen3-coder"
}

4) Run OpenCode

opencode

Notes

• Tool calls follow the normal OpenAI request/response loop.
• For Qwen coder models, --enforce-parser qwen_coder is usually the most reliable parser setting.
• If OpenCode reports a model mismatch, compare your configured model against GET /v1/models.

Troubleshooting

Chat logger:

export CANDLE_VLLM_CHAT_LOGGER=1

Reasoning routing for tool-enabled requests:

export CANDLE_VLLM_STREAM_AS_REASONING_CONTENT=1

Set CANDLE_VLLM_STREAM_AS_REASONING_CONTENT=0 if the client expects reasoning to stay inside content instead of separate reasoning_content fields.