OpenCode with a local Qwen3.6 35B-A3

17 avril 2026

Qwen just released a significant improvement to their models that are small enough to run on a standard (if high end) machine.

I’m test-running it on an apple M2 ultra (64GB of ram), and it’s surprisingly competent.

Making it analyse a recent commit, (done by claude Opus 3.6), it spotted a few potential issues, one was a red-herring it then spent significant time ensuring was actually fine, then it moved on with other fixes, that are valuable in themselves, with open code, it was able to work very independently.

To set it up, i downloaded the latest version of llama-cpp from github and extracted it (anywhere you like) then added that to my path (i use fish as a shell, so fish_add_path $PWD it is).

then ran it through.

 export LLAMA_CACHE="unsloth/Qwen3.6-35B-A3B-GGUF"
             llama-server \
                     -hf unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL \
                     --temp 0.6 \
                     --top-p 0.95 \
                     --top-k 20 \
                     --min-p 0.00 \
                     --ctx-size 1000000

it had to download it first, so, some patience warranted, as this is about 20GB of data.

Then it starts, and is accessible through http://localhost:8080

Good to play with, but the real deal is plugging it to opencode.

I had opencode setup previously, but i just had to update my configuration.

cat ~/.config/opencode/opencode.jsonc
{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "llama.cpp": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "llama-server (local)",
      "options": {
        "baseURL": "http://127.0.0.1:8080/v1"
      },
      "models": {
        "qwen3-coder:a3b": {
          "name": "Qwen3 (local)",
          "limit": {
            "context": 262144,
            "output": 65536
          }
        }
      }
    }
  },
  "tools": {
    "write": true,
    "bash": true,
    "webfetch": true
  }
}

Then it’s simply a matter of running opencode in the project directory, and start experimenting.