DashScope Qwen Native Model Retirement October 2026: The Routing Audit Your Team Cannot Skip

Alibaba Cloud Model Studio (Bailian) has published four separate retirement notices for October 10, 2026, covering a sweep of Qwen's own legacy model lineup — not just the third-party proxies. If your DashScope routing config still references qwen-turbo, qwq-plus, qvq-max, qwen-vl-max, or the paraformer ASR models, you have roughly 12 weeks before those model IDs stop returning responses. QPM throttling on the affected models has already started.

What Is Being Retired

Alibaba posted four official notices under a single October 10 sunset date. Across them, the following Qwen-native model IDs are going offline:

Language models:

• qwen-turbo and qwen-turbo-realtime — Qwen's previous general-purpose tier, replaced by current Qwen3.7/3.6 series endpoints
• qwen3.6-max-preview → qwen3.7-max
• qwen3-max-preview → qwen3-max
• qwen3-coder-plus → qwen3.7-plus
• qwen-coder-turbo, qwen-coder-plus (additional snapshot versions)
• qwq-plus — the QwQ reasoning model family
• qvq-max, qvq-plus — QVQ visual reasoning models
• qwen-math-turbo — no direct replacement named
• qwen3-max-2026-01-23, qwen3-max-2025-09-23 (dated snapshot variants)
• qwen3-vl-8b-instruct → qwen3.6-flash
• qwen-vl-max, qwen-vl-plus (prior-generation vision-language models)

Voice and audio:

• qwen-tts → cosyvoice-v3-flash
• qwen-tts-realtime
• qwen-voice-enrollment → voice-enrollment
• qwen-voice-design → cosyvoice-v3.5-flash
• gummy-chat-v1, gummy-realtime-v1 — voice translation models; no direct replacement listed for gummy-chat-v1

ASR (speech-to-text):

• paraformer-realtime-v1 → paraformer-realtime-v2
• paraformer-realtime-8k-v1 → paraformer-realtime-8k-v2
• paraformer-v1 → paraformer-v2
• paraformer-8k-v1 → paraformer-8k-v2
• paraformer-mtl-v1 → paraformer-v2

Why This Matters for AI Engineering Teams

This retirement wave is qualitatively different from the third-party Bailian proxy shutdowns announced earlier in July. Those affected teams using Bailian as a pass-through for DeepSeek, Kimi, or MiniMax. This one affects teams running Qwen's own production-grade inference APIs — including reasoning, vision-language, voice synthesis, and ASR.

Three operational implications stand out:

1. QPM throttling is already live. Per the official notice, Alibaba began reducing QPM (queries per minute) and TPM (tokens per minute) limits on affected models starting from the notice date — June 8, 2026 for the main lineup. By October 10, any call to a retired model ID will return an error. Teams that have been tracking elevated error rates or reduced throughput on qwen-turbo since mid-June should treat that as the migration pressure signal it is.

2. QwQ and QVQ are fully leaving DashScope. These reasoning and visual reasoning model names will stop responding on October 10. The replacement trajectory maps to qwen3.7-max (for reasoning-class workloads) and qwen3.6-flash (for vision tasks), but the model ID change is not backward-compatible — you cannot alias the old names. Any Claude Code, coding agent, or SDK configuration that hardcodes qwq-plus or qvq-max needs an explicit update.

3. Voice and ASR pipelines face a brand rename, not just a version bump. The qwen-tts → cosyvoice-v3-flash migration is a family rename, not a version increment. The cosyvoice endpoint likely has different API parameters, voice ID formats, and audio format options. Teams with voice agents or transcription pipelines should treat this as a compatibility break requiring explicit testing, not a model-string swap.

The Router/Operator Angle

From a routing policy standpoint, this event has three layers:

Model-ID audit. The highest-priority action is a full inventory of which DashScope model IDs are in use across services, CI pipelines, coding agents, and SDK configurations. Alibaba's own suggestion is to use the Model Telemetry console to see live usage by model ID under your account — this is the fastest way to identify undocumented or legacy configurations.

Replacement tier mapping. The new model IDs are not always obvious from the retirement table. Key mappings:

Fallback routing. For teams routing through a multi-provider gateway, the October 10 hard cutoff means any fallback chain that includes a retired Qwen model ID must be updated. A fallback to qwen-turbo after a primary deepseek-v4-flash failure will fail silently after October 10 unless the routing rule is updated to qwen3.7-flash. This is the scenario most likely to surface as an unexpected production incident — not the primary model changing, but a neglected fallback.

What TheRouter Users Should Watch

Teams routing Qwen models through TheRouter's DashScope provider lane should audit their routing tables for the retiring model IDs listed above. In particular:

• Review any route or fallback chain referencing qwen-turbo, qwq-plus, qvq-max, or qwen-vl-max.
• Check voice and ASR pipelines for qwen-tts, qwen-voice-enrollment, or paraformer-*-v1 patterns.
• Update fallback targets before October 10; do not wait for error-rate alerts.

For routing configuration reference, see the provider selection docs and the TheRouter docs for provider integration guides.

If gummy-chat-v1 (voice translation) is in active use and no replacement is available on DashScope, evaluate routing to an alternative translation-specialized API before the cutoff.

The notice period closes October 10. With QPM already throttled, any team still on these model IDs should treat the migration as overdue, not upcoming.