Use free Qwen3.8-27B in Claude Code with unoblox

Keep your familiar coding workflow. Add another model for the next small task.

If you already use Claude Code, trying another model should not mean rebuilding your development environment. You should be able to ask Qwen to explain a module or draft a focused change, then return to native Claude when you want a second review.

unoblox provides Qwen3.8-27B through an Anthropic-compatible gateway. Free access is currently available on unoblox; check the model page for current access details. Your native Claude usage continues under your existing Anthropic arrangement.

This guide uses Claude Code in the terminal, with Qwen and native Claude in separate sessions. It does not add Qwen to the Claude Desktop model picker. In our test of Claude Desktop for Mac 2.2553.1, Desktop explicitly removed Qwen from its gateway model list. Changing the model name does not switch providers or transfer your Claude subscription.

Start with a task you can check

Choose a small repository you know. A useful first task is explaining a module, writing a unit test, or documenting a function. You want a result you can inspect before asking for a larger change.

For example:

Read the repository instructions and inspect the request-validation code.
Do not edit anything yet.
Explain the validation rules, identify missing test cases, and list the files
that would need to change. Cite the relevant functions.

That gives you a concrete way to evaluate Qwen: did it read the right files, understand the rules, and propose relevant tests?

What you need

• Claude Code installed and working for your normal Claude sessions.
• Python 3 available as python3. The walkthrough was tested on macOS; Windows was not tested.
• An unoblox account with access to Qwen3.8-27B.
• An unoblox inference key. Follow the authentication guide.
• A project with a test command you can run.

The Claude Code integration guide explains the gateway connection. The endpoint origin is https://api.unoblox.ai; Claude Code adds /v1/messages itself.

Set up the Qwen session

Save the following script as claude-qwen.py in your home directory. It asks for your unoblox inference key without echoing it, starts Qwen with a separate user configuration directory, and leaves your normal Claude settings alone. It does not write the key to a configuration file.

#!/usr/bin/env python3
"""Start a separate unoblox Qwen session without changing native Claude settings."""
import getpass
import os
from pathlib import Path
import shutil
import subprocess
import sys


def main():
    claude = shutil.which('claude')
    if not claude:
        raise SystemExit('Install Claude Code and make sure claude is on PATH.')
    key = os.environ.get('UNOBLOX_API_KEY') or getpass.getpass('unoblox inference key: ')
    if not key.strip():
        raise SystemExit('An unoblox inference key is required.')
    config = Path.home() / '.config' / 'unoblox' / 'claude-code'
    config.mkdir(parents=True, exist_ok=True, mode=0o700)
    env = {k: v for k, v in os.environ.items()
           if not k.startswith('ANTHROPIC_') and k not in (
               'CLAUDE_CODE_USE_BEDROCK', 'CLAUDE_CODE_USE_VERTEX',
               'CLAUDE_CODE_USE_FOUNDRY')}
    env.update({
        'CLAUDE_CONFIG_DIR': str(config),
        'ANTHROPIC_BASE_URL': 'https://api.unoblox.ai',
        'ANTHROPIC_AUTH_TOKEN': key.strip(),
        'CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC': '1',
    })
    # Exclude repository-local settings that could override the gateway.
    # Organization-managed policies may still apply.
    return subprocess.call([
        claude, '--setting-sources', 'user',
        '--model', 'qwen/qwen3.8-27b', *sys.argv[1:]
    ], env=env)


if __name__ == '__main__':
    sys.exit(main())

Open a terminal in a small project you trust, then launch:

python3 ~/claude-qwen.py

Paste your unoblox inference key at the hidden prompt. Complete any first-run setup and project-trust prompts Claude Code presents. Keep normal tool-permission checks enabled. You can also supply the key through UNOBLOX_API_KEY using your existing secret-management tool; do not put a real key into a command you intend to share.

For a single read-only question:

python3 ~/claude-qwen.py -p --tools Read --allowedTools Read \
  "Read validation.py and explain its validation rules. Do not edit files."

Replace validation.py with a real file in your project. The launcher uses the exact catalog ID qwen/qwen3.8-27b and the gateway origin https://api.unoblox.ai.

The separate directory is ~/.config/unoblox/claude-code. The script excludes repository-local settings so they cannot redirect this session to another provider; repository-specific hooks, permissions and MCP settings may therefore need separate consideration. Organization-managed policies still apply. Keep this dedicated configuration free of conflicting gateway overrides. See Claude Code's environment reference for configuration behavior.

To return to your existing native Claude configuration, exit the Qwen session and run your usual command:

claude

The launcher changes only its child process environment, so its gateway credentials do not remain in your parent terminal. This assumes your normal Claude command was already configured for native Claude before starting.

What we checked on a real installation

We tested this connection using Claude Code 2.1.87 on macOS. In a disposable project, both native Claude and Qwen3.8-27B through unoblox read a Python validation function and correctly explained why it explicitly rejects booleans. The original Qwen run completed in about six seconds across two turns. We then ran the launcher printed above against the live gateway on 22 September 2026: it completed the same file-read task successfully in two turns, with no tool-permission denials. This was a small integration check, not a coding benchmark.

The Qwen session used separate user configuration and process-scoped credentials. The existing native Claude settings stayed unchanged, and native authentication remained available afterward. Session isolation is important: changing the gateway globally would affect normal Claude launches too.

Let Qwen work on one bounded change

Once the read-only response looks useful, ask for a narrowly defined implementation:

Add tests for the missing validation cases you identified.
Follow the repository's existing test framework and naming conventions.
Do not add dependencies or change production behavior.
Run the relevant tests and report the exact command and result.
If you cannot run them, say so explicitly.

Read the diff and run the tests yourself. A convincing explanation is not evidence that the implementation works.

For a documentation task, try:

Update the README example to match the current function signature.
Check every argument against the implementation.
Keep the change limited to documentation and show the diff.

These are suggested exercises, not benchmark results or claims that every task will succeed.

Bring the result back to native Claude

Finish the Qwen session, then open your normal native Claude session in the same repository. Hand over the actual files and evidence:

Review the current git diff without editing files.
Check whether the added tests cover real behavior and whether any assertions
could pass while the implementation is wrong.
Read the relevant implementation before drawing conclusions.
Report specific issues with file references, or say that none were found.

The repository is the handoff. A second session does not automatically know the first session's conversation. Include a short summary of the task, the changed files, the test command and any unresolved questions.

Avoid having two sessions edit the same files simultaneously. Sequential implementation and review keeps the handoff easier to understand.

Know which model is doing the work

Claude Code is the client. Qwen is the model serving the unoblox session. Keeping the same interface does not make the two models behave identically or guarantee every Claude-specific feature is available through the gateway.

unoblox's Messages endpoint translates function-tool calls and supports streaming. Its integration documentation also lists compatibility limits. If a request fails, check the selected catalog identifier, key access and error message before changing configuration.

A rate-limit response means you need to wait or reduce requests. A single coding task may generate several model requests as the agent reads files, calls tools and revises its answer. Free access does not imply an unlimited number of coding tasks.

Try one task, then decide where it fits

Start with one explanation, one test or one documentation change. Evaluate the diff, test results and amount of correction required. Keep the tasks that work well in your Qwen workflow, and use native Claude wherever you prefer its results.

Explore Qwen3.8-27B on unoblox.