Qwen3.6-Max vs Qwen3.7-Max: What Changed and When to Upgrade Your Routing Config

Update 2026-10-17: Alibaba retired qwen3-max, qwen3.6-max-preview, qwen3-coder-plus and qwen3-vl-flash on 2026-10-10, and TheRouter no longer serves them. Code examples below now use the live successors (qwen/qwen3.7-max, qwen/qwen3.7-plus, qwen/qwen3.7-flash).

Alibaba's DashScope deprecation page now lists qwen3.6-max-preview as sunsetting on September 8, 2026, with qwen3.7-max as the designated replacement. If you run any production workload against qwen3.6-max-preview, you have roughly 80 days to test the upgrade and switch your model ID — or lose API access entirely.

We dug into both models to compare what actually changed: context window, pricing structure, modality support, thinking mode behavior, and snapshot versioning. The short version: Qwen3.7-Max is cheaper per token for most workloads, supports a 4× larger context window, and adds a multimodal snapshot — but the pricing model shifted from tiered to flat-rate, and behavior differences in thinking mode may require testing before you switch.

OpenAI-compatible means a provider exposes a chat-completions endpoint whose request and response shape matches the OpenAI API contract closely enough that an unmodified OpenAI SDK call works against it after swapping three values: API key, base URL, and model name. The minimum surface in practice is POST /v1/chat/completions with messages, model, and an OpenAI-shaped streaming response.

TL;DR — Qwen3.6-Max-Preview vs Qwen3.7-Max at a Glance

Bottom line: For text-only workloads, Qwen3.7-Max is strictly better — larger context, lower output cost, batch API support. The only reason to delay switching is if you depend on Qwen3.6-Max-Preview's specific output behavior and need time to regression-test.

What Is Qwen3.6-Max-Preview?

Qwen3.6-Max-Preview launched on April 20, 2026 as Alibaba's then-flagship Max-tier model. It topped six agentic coding benchmarks at launch, including SWE-bench Pro and Terminal-bench. The "-Preview" suffix signaled active development — Alibaba explicitly stated the model was "still under active development" with "further gains in subsequent versions."

Key specs:

• Context window: 256K tokens with tiered pricing (two bands: ≤128K and 128K–256K)
• Pricing (China mainland): ¥9/¥15 input, ¥54/¥90 output per million tokens
• Modality: Text-only input and output
• Thinking mode: Supports both thinking and non-thinking modes
• Context cache: Supported with input token discount

The "Preview" label was the tell: this model was always meant to be superseded. Qwen3.7-Max is that successor.

What Is Qwen3.7-Max?

Qwen3.7-Max launched on May 20, 2026 as the next-generation Max-tier model. Alibaba describes it as "built to be a versatile agent foundation — equally capable of writing and debugging code, automating office workflows, and complex multi-step reasoning."

Two dated snapshots exist:

1. qwen3.7-max-2026-05-20 — Text-only. The launch snapshot, focused on coding, agentic tasks, and long-context reasoning. SWE-Bench Verified score of 80.4%.

2. qwen3.7-max-2026-06-08 — Multimodal. Added visual modality understanding on June 8, enabling image + text input. Alibaba's release notes describe it as having "multimodal interactive hybrid agent capabilities."

The alias qwen3.7-max currently points to the May 20 text-only snapshot. When you call qwen3.7-max without a date suffix, you get the text-only version.

Key specs:

• Context window: 1M tokens with flat-rate pricing
• Pricing (China mainland): ¥12 input, ¥36 output per million tokens (no tiers)
• Modality: Text-only (May 20) or text + image (June 8 snapshot)
• Thinking mode: Supports both thinking and non-thinking modes
• Batch API: Supported at 50% discount
• Context cache: Supported with input token discount

Pricing Comparison: Who Wins?

The pricing story depends on your context length.

Short context (≤128K tokens)

For short-context workloads, Qwen3.7-Max input costs more but output costs significantly less. Since most LLM API calls generate more value from output than input, Qwen3.7-Max is cheaper in practice for typical chat completions and code generation.

Medium context (128K–256K tokens)

At 128K+, Qwen3.7-Max wins on both input and output. The flat-rate model eliminates the tiered penalty entirely.

Long context (256K–1M tokens)

Qwen3.6-Max-Preview doesn't go here — its context window caps at 256K. Qwen3.7-Max handles up to 1M tokens at the same flat ¥12/¥36 rate. If you need long-context processing, there's no comparison.

International pricing

Pricing outside China mainland varies by region. Singapore runs roughly 1.5× the mainland price for both models. US (Virginia), Germany (Frankfurt), and Japan (Tokyo) regions serve Qwen3.7-Max at mainland rates for "global" deployments, but EU/Japan-specific endpoints cost more. Check the DashScope pricing page for your region.

USD approximations (at ¥7.3 = $1):

Two Snapshots: Which One Should You Use?

Qwen3.7-Max has two dated snapshots. This matters because they have different capabilities:

qwen3.7-max-2026-05-20 (text-only)

• Text input only
• SWE-Bench Verified 80.4%
• Agent-first design optimized for coding and complex multi-step tasks
• Best for: code generation, agentic workflows, long-context text processing

qwen3.7-max-2026-06-08 (multimodal)

• Text + image input
• Visual modality understanding for real-world scene perception
• Same pricing as the text-only snapshot
• Best for: image analysis, code-from-screenshot, document OCR, visual reasoning

If you don't need vision, pin to qwen3.7-max-2026-05-20 for deterministic behavior. If you need multimodal input, use qwen3.7-max-2026-06-08 explicitly. The unversioned qwen3.7-max alias currently points to the May 20 snapshot but will eventually roll forward.

Migration Checklist: Switching from Qwen3.6-Max-Preview

Step 1: Update the model ID

The only required change in your code:

from openai import OpenAI

client = OpenAI(
    api_key="your-dashscope-api-key",
    base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
)

# Before:
response = client.chat.completions.create(
    model="qwen3.6-max-preview",  # Sunsetting Sept 8
    messages=[{"role": "user", "content": "Explain quicksort"}],
)

# After:
response = client.chat.completions.create(
    model="qwen3.7-max",  # Active replacement
    messages=[{"role": "user", "content": "Explain quicksort"}],
)

Step 2: Test thinking mode behavior

Both models support thinking and non-thinking modes, but the internal reasoning may differ. If your application parses thinking tokens or relies on specific chain-of-thought patterns, run your evaluation suite against both models before switching.

Step 3: Check context length assumptions

If your application truncates input at 256K tokens (Qwen3.6-Max-Preview's limit), you can now increase that limit to 1M with Qwen3.7-Max. This is an opportunity, not a gotcha — but verify your token counting and chunking logic handles the larger window.

Step 4: Verify billing impact

Run a cost comparison on your actual traffic. For output-heavy workloads, you'll likely see savings. For input-heavy short-context workloads, the input price increase from ¥9 to ¥12 may slightly raise costs — but the output savings usually offset it.

Step 5: Test multimodal input handling (if using June 8 snapshot)

If you switch to qwen3.7-max-2026-06-08 for vision capabilities, test that your application handles the image input format correctly. DashScope uses the OpenAI-compatible image_url format:

response = client.chat.completions.create(
    model="qwen3.7-max-2026-06-08",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What's in this image?"},
            {"type": "image_url", "image_url": {"url": "https://example.com/image.png"}},
        ],
    }],
)

When to Stay on Qwen3.6-Max-Preview (Briefly)

There are a few cases where you might delay the switch:

1. Regression testing takes time. If your application relies on specific output patterns (e.g., structured JSON generation with tight format constraints), you may need a few weeks to validate Qwen3.7-Max produces equivalent results.

2. Budget planning for input cost increase. Short-context workloads see a 33% input price increase (¥9 → ¥12). If your workload is heavily input-bound with minimal output, run the numbers first.

3. Third-party integrations lag. Some platforms that host Qwen models (SiliconFlow, OpenRouter) may not immediately reflect the model ID change. Verify availability on your provider.

But don't wait too long — DashScope will gradually reduce QPM and TPM for sunsetting models starting from the deprecation notice date. Performance degradation begins before the final shutdown.

TheRouter Integration

If you route through TheRouter, updating your model ID in the routing configuration is the primary change. TheRouter supports OpenAI-compatible requests through configured providers, including Aliyun Bailian. Where fallback routing is configured, you can set qwen3.7-max as the primary model with qwen3.6-max-preview as a temporary fallback during your transition window:

# Example routing config during transition
primary: qwen/qwen3.7-max
# qwen3.6-max-preview was retired on 2026-10-10 — there is no fallback to keep

After validating Qwen3.7-Max output quality, remove the fallback to avoid routing to a model that will stop responding after September 8.

For model details, see the Qwen3.7-Max model page and the Qwen3-Max model page.

Timeline: What Happens When

Related Posts

• Qwen3.7-Max Multimodal vs GPT-4o vs Claude Sonnet 4 — Compares Qwen3.7-Max's multimodal snapshot against GPT-4o and Claude Sonnet 4.6 for vision + code workloads
• Aliyun Bailian API Guide — Complete DashScope setup: API keys, endpoints, model families, pricing tiers
• Qwen-Turbo, VL, Coder-Plus July 13 Sunset Migration Guide — Migration guide for the earlier July 13 deprecation wave
• Chinese AI Model Deprecation Calendar Q3 2026 — Full deprecation timeline across DashScope, SiliconFlow, and other Chinese providers
• OpenAI-Compatible API Providers — Overview of all OpenAI-compatible providers including DashScope

---

Sources:

• DashScope Model Deprecation Page — Retrieved June 19, 2026
• DashScope Model Pricing — Retrieved June 19, 2026
• DashScope Newly Released Models — Retrieved June 19, 2026
• Qwen3.7 Blog Post — Retrieved June 19, 2026
• Qwen3.6-Max-Preview Blog Post — Retrieved June 19, 2026
• Aliyun Model Studio Supported Models — Retrieved June 19, 2026