Alibaba Bailian Retires DeepSeek, Kimi, MiniMax, and GLM Models on October 10: The Multi-Provider DashScope Routing Audit

Alibaba Cloud Model Studio (Bailian) has published its deprecation schedule for October 10, 2026, and the scope is wider than most teams realize. Beyond retiring legacy Qwen variants, Bailian is pulling the plug on every third-party model it has been proxying through DashScope's OpenAI-compatible endpoint — including DeepSeek V3, DeepSeek R1, Kimi K2, MiniMax M2.1, and both GLM 4 versions from Zhipu. If your team has been using Bailian as a unified API hub for domestic Chinese models, October 10 is now your hard migration deadline.

What Is Bailian's Role as a Multi-Model Proxy?

DashScope's OpenAI-compatible endpoint (https://dashscope.aliyuncs.com/compatible-mode/v1) is not just a gateway to Qwen models. Alibaba Cloud Model Studio has been hosting third-party models from DeepSeek, Moonshot, MiniMax, Zhipu, and others under the same API surface — letting teams hit a single endpoint with a single key to access models from multiple domestic vendors. The tradeoff: those model IDs are now on a retirement schedule dictated by Alibaba, not the original vendors.

The Full October 10 Retirement List

Third-Party Models (Complete Removal)

Alibaba is retiring all the following model IDs from DashScope as of October 10, 2026:

DeepSeek:

• deepseek-v3, deepseek-v3.1, deepseek-v3.2, deepseek-v3.2-exp
• deepseek-r1, deepseek-r1-0528
• deepseek-r1-distill-qwen-7b, deepseek-r1-distill-qwen-14b, deepseek-r1-distill-qwen-32b

Moonshot / Kimi:

• Moonshot-Kimi-K2-Instruct, kimi-k2

MiniMax:

• MiniMax-M2.1

Zhipu / GLM:

• glm-4.7, glm-4.6

Qwen First-Party Models Also Retiring

The same October 10 deadline covers a broad sweep of Qwen first-party models:

Qwen Coder:

• qwen3-coder-next, qwen3-coder-plus (main line) → both replaced by qwen3.7-plus
• qwen3-coder-plus-2025-09-23, qwen3-coder-plus-2025-07-22 (snapshots)
• qwen3-coder-30b-a3b-instruct, qwen3-coder-480b-a35b-instruct
• qwen-coder-turbo, qwen-coder-plus

Qwen Turbo / Reasoning:

• qwen-turbo, qwen-turbo-realtime
• qwq-plus, qvq-max, qvq-plus
• qwen-math-turbo

Qwen3 Open-Source Variants:

• qwen3-8b, qwen3-14b, qwen3-32b, qwen3-30b-a3b, qwen3-235b-a22b
• qwen3-30b-a3b-instruct-2507, qwen3-30b-a3b-thinking-2507
• qwen3-235b-a22b-instruct-2507, qwen3-235b-a22b-thinking-2507
• qwen3-next-80b-a3b-instruct, qwen3-next-80b-a3b-thinking

Qwen VL:

• qwen3-vl-flash, qwen-vl-max, qwen-vl-plus
• Multiple qwen3-vl-* snapshot and size variants

Speech Models:

• qwen-tts, qwen-tts-realtime, qwen-voice-enrollment, qwen-voice-design
• gummy-chat-v1, gummy-realtime-v1
• Multiple Paraformer speech recognition models

The Throttling Timeline Starts Now

Alibaba's deprecation policy has an important detail: QPM (queries per minute) and TPM (tokens per minute) limits begin to shrink from the notice date, not just at shutdown. If you applied for quota increases on any of these models, those are already being rolled back to default limits. The API interface remains available until October 10, but capacity is narrowing progressively.

This means teams who are heavy users of deepseek-v3 or qwen-coder-plus via DashScope may already be hitting lower rate limits than expected.

Why This Matters for Routing Layer Design

The retirement of all third-party model proxies from a single vendor reveals a structural risk in multi-model routing:

Vendor-specific model IDs tie you to a single upstream. A team using deepseek-v3 via DashScope cannot port that config directly to DeepSeek's own API — the model ID is the same, but the base URL, authentication format, and quota management differ. Any gateway or routing layer that stores deepseek-v3 as a model name needs to know which upstream that resolves to.

Proxy catalogs change on the proxy vendor's schedule. The DeepSeek API is not retiring deepseek-v3 on October 10 — Alibaba is retiring their proxy copy. Teams who assumed DashScope's third-party coverage would track vendor releases indefinitely now have a hard deadline.

The migration target for most third-party models is the original vendor's API. For deepseek-v3 → connect directly to api.deepseek.com. For kimi-k2 → use Moonshot's API. For glm-4.6/4.7 → use Zhipu's BigModel API. For MiniMax-M2.1 → use MiniMax's API directly. DashScope's own replacement recommendations for third-party models are sparse because the answer is: go back to the source.

The Routing Audit Checklist

Before October 10, every team routing traffic through DashScope should:

1. Check active model usage — Visit Bailian Model Telemetry to see which model IDs your account has called in the past 30 days.

2. Identify which are retiring — Cross-reference against the full deprecation list at help.aliyun.com/zh/model-studio/model-depreciation.

3. Plan the migration target for each third-party model:

   • DeepSeek V3/R1 → DeepSeek API directly (api.deepseek.com)
   • Kimi K2 → Moonshot AI API (api.moonshot.cn)
   • MiniMax M2.1 → MiniMax API (api.minimax.chat)
   • GLM 4.6/4.7 → Zhipu BigModel API (open.bigmodel.cn)

4. For Qwen first-party → use the recommended replacements:

   • qwen-coder-plus / qwen3-coder-plus → qwen3.7-plus
   • qwen-turbo → latest Qwen3.7/3.6 series
   • qwen3-8b, qwen3-14b → qwen3.6-flash
   • qwen3-30b-a3b → qwen3.7-plus

5. Update routing configs, not just application code — Model names in gateway configs, prompt templates, tool-call formatters, and token budget calculators all need to change.

6. Test replacements under load before the QPM throttle becomes a problem — The window to test qwen3.7-plus as a drop-in for qwen3-coder-plus is now, while you still have full quota on both.

The Official Source

The full deprecation page is maintained at: https://help.aliyun.com/zh/model-studio/model-depreciation

Supporting notices are at aliyun.com/notice/118344, /118345, /118177, /118331. The page lists each retiring model ID, its replacement, and the notice date — the most complete reference available.

Teams that built multi-model access on Alibaba's DashScope proxy layer got a compelling deal: one key, one endpoint, a dozen domestic models. October 10 is the price of that convenience. The migration path is straightforward — but only if you know every model ID you're calling and where each one needs to go next.