Alibaba ships Qwen models at a pace that makes it hard to keep up. Between Qwen3, Qwen3 Coder, Qwen3.5, and the A-variants, there are a dozen models that could plausibly run in OpenClaw. Most of them aren’t worth the setup time.
Short version: pick one of three depending on what you’re doing.
The quick answer
| Model | Input/Output Cost | Context | Best For |
|---|---|---|---|
| Qwen3 Coder | $0.22 / $1.00 | 262K | Default for coding |
| Qwen3 Coder Plus | $0.65 / $3.25 | 1M | Harder problems, long context |
| Qwen3.5-Flash | $0.07 / $0.26 | 1M | Cheap fast path, large context |
| Qwen3.5 397B A17B | $0.39 / $2.34 | 262K | Reasoning + vision |
| Qwen2.5 Coder 32B | $0.66 / $1.00 | 32K | Legacy — use Qwen3 Coder |
Most people should start with Qwen3 Coder and only reach for Coder Plus when the task actually needs it.
Qwen3 Coder — the default pick
$0.22/M input, $1/M output, 262K context. Qwen3 Coder is Alibaba’s dedicated coding model and it punches well above its price tag. On day-to-day OpenClaw workflows (refactoring, file edits, bug fixing, explaining code), it holds up against models that cost 5–10x more.
Function calling works. The model is reasonably good at generating clean diffs, and the 262K context window is enough to load a substantial codebase or multiple large files. Where it struggles is novel algorithm design and genuinely hard debugging. For those, you want Coder Plus or a Western flagship.
For interactive OpenClaw sessions where cost matters, this is a solid default. It’s particularly good for high-frequency workflows where you’re making many small edits, since the low output token cost keeps the bill in check.
Qwen3 Coder Plus — the step-up
$0.65/M input, $3.25/M output, 1M token context. Coder Plus is what you reach for when regular Coder isn’t enough. The quality jump is real on hard problems, and the 1M context window lets you load whole monorepos.
Reasoning is supported, which helps on multi-step tasks. The model is noticeably more reliable on complex refactors that span many files. Regular Coder can lose the thread; Coder Plus usually doesn’t.
Still cheaper than most Western flagships. At $0.65/$3.25, it’s about a fifth the cost of Claude Sonnet 4.6 and roughly a quarter the cost of GPT-5.4. For teams trying to cut their OpenClaw bill without dropping into budget-model territory, Coder Plus is the sweet spot.
Qwen3.5-Flash — the ultra-cheap fast path
$0.07/M input, $0.26/M output, 1M context. Qwen3.5-Flash is absurdly cheap for what it does. It’s not a coding-specialized model, so it won’t match Qwen3 Coder on hard programming tasks, but it’s capable at general reasoning, summarization, and long-document work.
Use this when you need a large context window on a budget: log analysis, doc generation, summarizing PR changes, generating commit messages at scale. For interactive coding, stick with Qwen3 Coder.
Qwen3.5 397B A17B — the reasoning + vision option
$0.39/M input, $2.34/M output, 262K context. The A17B variant supports reasoning and vision, which the Coder models don’t. If OpenClaw needs to look at screenshots, read UI mockups, or do chain-of-thought reasoning on non-code problems, this is the Qwen model to pick.
Cost-wise it sits between Qwen3 Coder and Coder Plus. It’s not a coding specialist, so for pure code work Qwen3 Coder Plus is the better choice. But for mixed workflows that involve vision or hard reasoning alongside code, A17B covers more ground.
Legacy models to skip
- Qwen2.5 Coder 32B ($0.66/$1.00): Superseded by Qwen3 Coder at a lower price with a much bigger context window. No reason to use it.
- Qwen3 235B A22B / A22B 2507: These are general-purpose, not coding-tuned. Qwen3 Coder is better at the work OpenClaw actually does.
Setup in OpenClaw
The fastest path to Qwen in OpenClaw is through haimaker.ai. Qwen models are available alongside Claude, GPT, Gemini, and dozens of other models through a single API key. No separate Alibaba Cloud account needed.
1. Get your haimaker.ai API key
Sign up at haimaker.ai and copy your key from the dashboard.
2. Add haimaker as a provider
Open ~/.openclaw/openclaw.json:
{
"models": {
"providers": {
"haimaker": {
"baseUrl": "https://api.haimaker.ai/v1",
"apiKey": "your-haimaker-api-key",
"api": "openai-completions"
}
}
}
}
3. Add Qwen models to the allowlist
{
"agents": {
"defaults": {
"models": {
"haimaker/qwen/qwen3-coder": {},
"haimaker/qwen/qwen3-coder-plus": {},
"haimaker/qwen/qwen3.5-flash": {}
}
}
}
}
4. Apply the config
Run openclaw gateway config.apply and switch models with /model during a session.
What I’d do
Default to Qwen3 Coder. Step up to Coder Plus when the task is genuinely hard or you need the 1M context window. Drop to Qwen3.5-Flash when output token cost is the bottleneck and you’re doing batch work.
If you’re running OpenClaw on a tight budget, the Qwen lineup gets you 80% of the capability of a Western flagship at 10–20% of the cost. It’s not as strong as Sonnet 4.6 or GPT-5.4 on the hardest problems, but for the stuff OpenClaw actually spends most of its time on, the gap is much smaller than the price difference.