feat: Add Alibaba Cloud / Qwen Models as a Provider

[chindris-mihai-alexandru]

Feature Request: Add Alibaba Cloud / Qwen Models as a Provider

Summary

Add Alibaba Cloud's Qwen models as a first-class provider in AdaL CLI, including:

1. BYOAK (Bring Your Own API Key) support for DashScope API keys
2. Qwen cloud models in the /model selector (Qwen3 Coder Plus, Qwen3 Coder Flash, Qwen3.5 Plus)
3. Partnership integration (optional) — offer Qwen models through AdaL subscription tiers, similar to the existing ZAI/GLM partnership

Why This Matters

1. Competitive Gap

AdaL currently supports cloud models from OpenAI, Anthropic, Google, ZAI, and MiniMax — but not Alibaba Cloud / Qwen via cloud API, despite Qwen3 Coder being one of the strongest open-weight coding models available. (AdaL does support qwen3-coder:30b locally via Ollama, but cloud API access to the full-sized models is missing.)

Other coding agents are already adding Alibaba Cloud support:

• ForgeCode (6,600+ stars) added AlibabaCoding as a built-in provider with Qwen3.5+, Qwen3 Coder, and more (PR #2696)
• Pi Coding Agent has a working Qwen CLI extension with OAuth-based free access and full thinkingFormat: "qwen" support baked into the core (PR #940)

2. Alibaba Cloud Pricing Changes Create an Opportunity

Alibaba Cloud recently removed their $3/month Lite plan, leaving only $50+/month plans. Individual developers who want occasional Qwen API access are now priced out of direct DashScope access.

AdaL is perfectly positioned to fill this gap — just as it already does with ZAI/GLM models. By offering Qwen through AdaL's $20–$200/month subscription tiers, users get cost-effective access to Qwen models without needing a $50 Alibaba Cloud subscription.

3. Strong Model Lineup

Model	Context	Strengths
Qwen3 Coder Plus	1M tokens	Coding-optimized, massive context window
Qwen3 Coder Flash	1M tokens	Fast, budget coding tasks
Qwen3.5 Plus	256K tokens	Multimodal (text + image), hybrid thinking, 119+ languages

The 1M context window on Qwen3 Coder models matches Gemini and exceeds Claude (200K) and GPT (272K) — a real differentiator for large codebase work.

4. Existing Partnership Model Proves the Path

AdaL already has a successful partnership with ZhipuAI (Z.AI) offering GLM-5, GLM-4.7 FlashX, etc. as first-class models. Notably, Alibaba Cloud's DashScope platform also hosts GLM and Kimi models alongside Qwen — so a single Alibaba Cloud integration could potentially expand the model catalog further.

Proposed Implementation

Phase 1: BYOAK Support (Minimal Effort)

DashScope speaks the OpenAI-compatible API (dashscope.aliyuncs.com/compatible-mode/v1), so integration is straightforward:

• Add DASHSCOPE_API_KEY to /byoak alongside Anthropic, OpenAI, and Google
• Register Qwen models in the model selector under a "Qwen" or "Alibaba Cloud" provider section
• Add enable_thinking parameter support for Qwen's reasoning mode

Phase 2: Subscription Integration (Partnership)

• Negotiate with Alibaba Cloud to offer Qwen models through AdaL credits (similar to ZAI partnership)
• Add Qwen models to the Recommended section in /model
• Include Qwen in billing/pricing documentation

Phase 3: Free Tier via chat.qwen.ai OAuth (Optional)

Alibaba offers free API access (2,000 requests/day) through chat.qwen.ai OAuth with device code flow + PKCE — no Alibaba Cloud subscription required. This is the same model as how ChatGPT Subscription works in AdaL today, or how Google Gemini CLI provides free access with a Google account. Pi Coding Agent already confirmed this works (issue #919: "Qwen CLI provides capable models for free with generous limits"). AdaL could offer this as a "Qwen Free" option in /model → Third Party Subscriptions.

Technical Notes

• API compatibility: DashScope uses OpenAI Chat Completions format — same as what AdaL already supports
• Thinking/reasoning: Qwen uses a top-level enable_thinking: true parameter (not OpenAI's reasoning_effort), which needs a small adapter
• Known issues: ForgeCode's Alibaba integration hit 404 errors on their Anthropic-compatible endpoint (#2781) — recommend using the OpenAI-compatible endpoint (/compatible-mode/v1) instead, which is more stable
• Prompt caching: DashScope supports Anthropic-style cache_control for prompt caching (Pi is actively adding this: #3392)

References

• DashScope API docs
• Qwen3 Coder announcement
• chat.qwen.ai — free consumer access with OAuth
• Pi's Qwen CLI extension — reference implementation
• ForgeCode AlibabaCoding provider
• AdaL BYOAK docs