Ollama Qwen Thinking

An OpenCode plugin + local proxy that makes Qwen3.8 reasoning levels work transparently with Ollama.

OpenCode's Ollama integration uses the OpenAI-compatible /v1/chat/completions endpoint. OpenCode 1.18.x does not reliably forward arbitrary reasoning_effort request-body fields through the AI SDK. This project solves that at the HTTP boundary: the plugin starts a small localhost proxy and points the Ollama provider at it; the proxy rewrites Qwen3.8 model-name suffixes into Ollama's reasoning_effort field.

The problem this project solves

Qwen3.8 supports different reasoning levels such as none, low, medium, and xhigh. Ollama's OpenAI-compatible API can express the desired level through the reasoning_effort request parameter.

The problem is that when Qwen3.8 is used through OpenCode → Ollama, the reasoning-level parameter does not reliably make it all the way from OpenCode to Ollama. As a result, simply defining or selecting different reasoning levels in OpenCode may still produce the same effective reasoning behavior.

This project works around that request-path limitation without requiring changes to OpenCode or Ollama:

1. The plugin starts a small local HTTP proxy automatically.
2. The proxy exposes reasoning-level variants of each Qwen3.8 model, using IDs such as qwen3.8:27b-mlx-effort-medium.
3. When OpenCode requests one of those variants, the proxy converts the suffix into Ollama's reasoning_effort parameter.
4. Ollama receives the original model ID plus the requested reasoning level.

In short:

OpenCode
   |
   | qwen3.8:27b-mlx-effort-medium
   v
Qwen Thinking Proxy
   |
   | model=qwen3.8:27b-mlx
   | reasoning_effort=medium
   v
Ollama

The result is that the four reasoning levels can be exposed as ordinary selectable OpenCode models, while the actual reasoning-level translation happens at the HTTP boundary.

Supported models

The proxy is model-agnostic and will rewrite any Qwen3.8 model:

• qwen3.8:27b-mlx
• qwen3.8:27b-mtp-q4_K_M

and exposes reasoning variants:

• qwen3.8:27b-mlx-effort-none
• qwen3.8:27b-mlx-effort-low
• qwen3.8:27b-mlx-effort-medium
• qwen3.8:27b-mlx-effort-xhigh
• qwen3.8:27b-mtp-q4_K_M-effort-none
• qwen3.8:27b-mtp-q4_K_M-effort-low
• qwen3.8:27b-mtp-q4_K_M-effort-medium
• qwen3.8:27b-mtp-q4_K_M-effort-xhigh

The unsuffixed model is passed through unchanged.

Architecture

OpenCode
   |
   | OpenAI-compatible request
   v
Qwen Thinking Proxy :11437
   |
   | qwen3.8:27b-mlx-effort-medium
   |        -> model=qwen3.8:27b-mlx
   |        -> reasoning_effort=medium
   v
Ollama :11434

The proxy is started automatically by the OpenCode plugin. No separate service manager is required.

Prerequisites

Before installing the plugin, make sure the following are available on your machine:

• OpenCode 1.18.x or later with plugin support.
• Ollama installed and running locally.
• An Ollama-supported Qwen3.8 model, such as qwen3.8:27b-mlx or a Qwen3.8 MTP variant.
• Bun 1.1.26 or later. The plugin uses Bun to start the local proxy as a child process, and the configuration-generator script is also written for Bun. The proxy disables Bun.serve's idle timeout so long generations are not cut off, and that option requires Bun 1.1.26; on older versions it is silently ignored.
• Git, if installing directly from this GitHub repository.
• A standard JSON-formatted OpenCode configuration if you plan to use the generate:variants utility. The generator does not currently parse JSONC comments or trailing commas.

Verify the prerequisites

Check Bun:

bun --version

Check Ollama:

ollama --version
ollama list

Check that Ollama's OpenAI-compatible endpoint is available:

curl http://127.0.0.1:11434/v1/models

You should see your installed models in the response.

For the Qwen reasoning proxy, the important requirement is that the underlying Qwen3.8 model is available through Ollama. The plugin does not download or manage models itself.

Installation

npm package (recommended)

Once published, the easiest installation is to add the package directly to your OpenCode configuration:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": [
    "opencode-ollama-qwen-thinking@0.1.2"
  ]
}

You do not need to clone the repository or run npm install manually. OpenCode installs npm plugins automatically using Bun and caches them locally. citeturn0search0turn0search2

If you prefer to always use the latest published version, you can omit the version:

{
  "plugin": [
    "opencode-ollama-qwen-thinking"
  ]
}

For reproducible setups, pinning a specific version is recommended.

From this repository (development)

Clone the repository when you want to inspect, modify, or develop the plugin locally:

git clone https://github.com/amit-batra/opencode-ollama-qwen-thinking.git

Then add the local repository path to your OpenCode configuration:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": [
    "/absolute/path/to/opencode-ollama-qwen-thinking"
  ]
}

Using the models

After restarting OpenCode, select one of the following model IDs:

ollama/qwen3.8:27b-mlx-effort-none
ollama/qwen3.8:27b-mlx-effort-low
ollama/qwen3.8:27b-mlx-effort-medium
ollama/qwen3.8:27b-mlx-effort-xhigh

ollama/qwen3.8:27b-mtp-q4_K_M-effort-none
ollama/qwen3.8:27b-mtp-q4_K_M-effort-low
ollama/qwen3.8:27b-mtp-q4_K_M-effort-medium
ollama/qwen3.8:27b-mtp-q4_K_M-effort-xhigh

The original model IDs remain valid:

ollama/qwen3.8:27b-mlx
ollama/qwen3.8:27b-mtp-q4_K_M

For an unsuffixed model, the request is forwarded without a reasoning_effort override, preserving Ollama's default behavior.

Generate Qwen3.8 model entries from an existing OpenCode config

If you already have one Qwen3.8 model entry in opencode.json, the repository includes a small Bun/TypeScript utility that creates four separate model entries for the proxy's reasoning levels:

• <base-model>-effort-none
• <base-model>-effort-low
• <base-model>-effort-medium
• <base-model>-effort-xhigh

The script copies the original model configuration, changes the model ID, and gives each generated entry a descriptive name. If the source entry has an explicit modelID, that field is updated too, so OpenCode sends the suffixed ID to the proxy. OpenCode's provider model map uses these model IDs as selectable catalog entries. citeturn0search0turn0search3

Automatic detection

If your provider contains exactly one unsuffixed Qwen3.8 model, the script can detect it automatically:

bun run generate:variants -- --input ~/.config/opencode/opencode.json

This creates:

~/.config/opencode/opencode.json.qwen-thinking.json

Specify the model explicitly

This is recommended if you have more than one Qwen3.8 model:

bun run generate:variants -- \
  --input ~/.config/opencode/opencode.json \
  --model qwen3.8:27b-mlx \
  --output ~/opencode-qwen-thinking.json

You can also select a different provider:

bun run generate:variants -- \
  --input ./opencode.json \
  --provider ollama \
  --model qwen3.8:27b-mlx

Update the existing file in place

Use --in-place when you want the four generated entries written directly into the original file:

bun run generate:variants -- \
  --input ~/.config/opencode/opencode.json \
  --model qwen3.8:27b-mlx \
  --in-place

The script does not overwrite an existing generated variant unless --force is supplied:

bun run generate:variants -- \
  --input ~/.config/opencode/opencode.json \
  --model qwen3.8:27b-mlx \
  --in-place \
  --force

Before and after

Given this source entry:

{
  "provider": {
    "ollama": {
      "models": {
        "qwen3.8:27b-mlx": {
          "name": "Qwen3.8 27B MLX",
          "options": {
            "temperature": 1,
            "num_ctx": 131072
          }
        }
      }
    }
  }
}

the generated config contains the original entry plus four entries with the same settings:

qwen3.8:27b-mlx-effort-none   -> No Thinking
qwen3.8:27b-mlx-effort-low    -> Low Thinking
qwen3.8:27b-mlx-effort-medium -> Medium Thinking
qwen3.8:27b-mlx-effort-xhigh  -> Xhigh Thinking

The generated entries deliberately do not add a separate reasoning parameter. The proxy selects the reasoning level from the -effort-<level> suffix and converts it into Ollama's reasoning_effort request field.

The utility currently accepts standard JSON files. If your config uses JSON comments or trailing commas, remove those first or save a JSON copy for the generator.

Environment variables

None of these environment variables are mandatory. The plugin and proxy have sensible defaults, so a normal installation requires no environment-variable configuration.

You only need to define an environment variable when you want to override its default behavior:

Variable	Default	Mandatory?	Purpose
OLLAMA_URL	http://127.0.0.1:11434	No	URL of the Ollama server
QWEN_THINKING_PROXY_HOST	127.0.0.1	No	Local address on which the proxy listens
QWEN_THINKING_PROXY_PORT	11437	No	Local TCP port used by the proxy
QWEN_THINKING_DEBUG	0	No	Set to 1 to log model/reasoning rewrites
QWEN_THINKING_DISABLE_PROXY	0	No	Set to 1 to prevent the plugin from starting/configuring the proxy automatically

Default installation

For the normal setup — Ollama running on the same machine at its default port — you do not need to define anything. Simply install the plugin and start OpenCode.

The effective defaults are:

Ollama:       http://127.0.0.1:11434
Proxy host:   127.0.0.1
Proxy port:   11437
Debug logging: disabled
Auto-start:   enabled

When would you override them?

For example, if Ollama is running on another port:

OLLAMA_URL=http://127.0.0.1:11435 opencode

If you want the proxy on a different port:

QWEN_THINKING_PROXY_PORT=12437 opencode

If you want request-rewrite logging:

QWEN_THINKING_DEBUG=1 opencode

Note: QWEN_THINKING_DISABLE_PROXY=1 is an advanced option. With it enabled, the plugin deliberately does not start the proxy or redirect the Ollama provider to it, so the -effort-* model variants will not provide the reasoning-level translation described by this project.

Configuration

Environment variables:

Variable	Default	Purpose
OLLAMA_URL	http://127.0.0.1:11434	Ollama upstream
QWEN_THINKING_PROXY_HOST	127.0.0.1	Proxy bind address
QWEN_THINKING_PROXY_PORT	11437	Proxy port
QWEN_THINKING_DEBUG	0	Log request rewrites
QWEN_THINKING_DISABLE_PROXY	0	Do not auto-start the proxy

For example:

QWEN_THINKING_DEBUG=1 opencode

How it works

A request for:

qwen3.8:27b-mlx-effort-medium

is transformed into:

{
  "model": "qwen3.8:27b-mlx",
  "reasoning_effort": "medium"
}

and forwarded to:

http://127.0.0.1:11434/v1/chat/completions

The proxy does not alter messages, tools, sampling parameters, streaming, or response bodies.

Streaming and long generations

The proxy disables Bun.serve's idle timeout. By default Bun closes a connection after 10 seconds without data, and that timer also runs while a response is being streamed, so on a local model it can reset the connection in the middle of an answer — for example while Qwen is still thinking at xhigh, or while the model is being loaded into memory. Long pauses between tokens are normal and are not interrupted.

The proxy listens on 127.0.0.1 only, and it already cancels the Ollama request when OpenCode disconnects, so disabling the timeout does not leave work running unnoticed.

For none, the proxy sends reasoning_effort: "none". This is important for Qwen3.8 because recent Ollama reports show that think: false can be intermittent through the OpenAI-compatible endpoint while reasoning_effort: "none" is reliable.

Health check

Once OpenCode has started the plugin:

curl http://127.0.0.1:11437/health

Expected:

{"ok":true,"service":"opencode-ollama-qwen-thinking","pid":1234,"upstream":"http://127.0.0.1:11434","idleTimeout":0}

idleTimeout should be 0. See Debugging if it is not.

List models through the proxy:

curl -s http://127.0.0.1:11437/v1/models | jq

Debugging

Enable request logging:

QWEN_THINKING_DEBUG=1 opencode

You should see entries similar to:

[proxy] qwen3.8:27b-mlx-effort-medium
        -> qwen3.8:27b-mlx
        reasoning_effort=medium

This gives a direct way to verify that the thinking level is actually reaching Ollama.

"Connection reset by server" or "Bun.serve() timed out a request after 10 seconds"

If the proxy logs

warn: Bun.serve() timed out a request after 10 seconds. Pass idleTimeout to configure.

then the running proxy is too old to disable that timer, and it is closing the connection whenever a generation pauses for more than 10 seconds. OpenCode reports this as Error: Connection reset by server and retries, hitting the same limit each time.

This was fixed in 0.1.2. Upgrade the plugin and restart OpenCode so the new proxy process starts. You can confirm the running version is current by checking that /health reports an idleTimeout of 0:

curl http://127.0.0.1:11437/health

Expected:

{"ok":true,"service":"opencode-ollama-qwen-thinking","pid":1234,"upstream":"http://127.0.0.1:11434","idleTimeout":0}

Requirements

• macOS or Linux
• OpenCode 1.18.x
• Ollama with an OpenAI-compatible /v1 endpoint
• Bun (OpenCode itself uses Bun for its plugin runtime)
• Qwen3.8 model support in your Ollama version

Why a proxy?

OpenCode plugins expose chat.params, but that hook is intended for parameters such as temperature, top-p, and provider options. The OpenAI-compatible provider controls which request-body fields ultimately reach Ollama. A localhost HTTP proxy operates after that filtering, so it can reliably rewrite the actual request sent to Ollama.

This is also why the plugin and proxy live in one repository: the plugin is just the lifecycle/bootstrapping layer; the proxy contains the protocol logic.

Development

Install dependencies:

bun install

Run the proxy directly:

bun run src/proxy.ts

Run tests:

bun test

License

MIT