Qwen3.8-Max vs DeepSeek V4 Pro: Chinese Flagship Model API Comparison for Routing Operators

If you are building on Chinese LLM APIs in August 2026, two models sit at the top of the stack: Alibaba's Qwen3.8-Max and DeepSeek's V4 Pro. Both are trillion-scale Mixture-of-Experts models with million-token context windows and OpenAI-compatible endpoints. Both support thinking mode. Both are available through DashScope as well as through their own direct APIs. The question is not whether they are good — they are — but which one fits your workload, and whether you should route between both.

We tested them across our own routing layer, compared the official specs and pricing, and put this guide together for operators who need to choose one or run both behind a single base_url.

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 Comparison

The short version: Qwen3.8-Max is the model to pick when you need vision or multimodal input alongside text. DeepSeek V4 Pro is the model to pick when your workload is text-only and you want the lowest per-token cost among Chinese flagships.

Architecture: Two Approaches to Trillion-Scale MoE

Both models use sparse Mixture-of-Experts architectures, but the designs differ in important ways.

Qwen3.8-Max is Alibaba's largest model, announced on July 19, 2026, and released via DashScope on August 3. It has roughly 2.4 trillion total parameters. Alibaba has not published the MoE configuration — we do not know the expert count, the number of active parameters per token, or the routing strategy. What we do know: it accepts text, image, and video input natively, supports a 1M-token context window, and offers hybrid reasoning (thinking mode toggled per request via standard parameters).

DeepSeek V4 Pro launched as a preview on April 24, 2026, with the official GA and peak/off-peak pricing arriving in mid-July. It has 1.6 trillion total parameters with 49 billion active per token — a published, concrete number. It is text-only, supports 1M-token context, and enables thinking mode by default (you can turn it off per request).

The practical difference: Qwen3.8-Max has more total parameters and native multimodal input. DeepSeek V4 Pro has a smaller active parameter count, which translates to faster inference and lower cost per token.

Pricing: DashScope vs Direct API

Pricing is the single biggest differentiator for routing operators choosing between these two models.

DashScope (Alibaba Cloud Bailian) Pricing

Both models are available on DashScope through the same OpenAI-compatible endpoint. Prices are in CNY per million tokens:

DeepSeek Direct API Pricing

DeepSeek V4 Pro is also available through DeepSeek's own API at https://api.deepseek.com:

DeepSeek's direct API pricing is denominated in USD and sits roughly in the same range as the DashScope CNY pricing after conversion. The cache-hit discount on DeepSeek's direct API is aggressive — $0.003625 per million input tokens is essentially free for repeat prompts.

For routing operators: if you are already routing through DashScope for Qwen models, adding DeepSeek V4 Pro on the same endpoint means no extra integration work. If you want the absolute lowest cost, DeepSeek's direct API with cache hits is hard to beat.

Note: DeepSeek has announced plans to raise their API pricing, with "a significant increase expected." We will update this comparison when the new pricing takes effect.

When comparing API pricing across providers, always normalize to USD per million tokens and split input from output. Most providers price output tokens 2–5× higher than input tokens, so a workload heavy on completion length looks very different from a retrieval-heavy workload at the same nominal "price per million."

• Use one currency (USD) — convert at publish date and cite the rate.
• Split input/output — never quote a single blended number.
• Cite each row to the provider's own pricing page with retrieval date.
• Note context-window tiers — long-context pricing often steps higher.

Context Window and Output Limits

Both models support 1M-token context windows, putting them at the top of the context-length chart alongside models like Gemini 2.5 Pro.

The output limit is where they diverge:

• Qwen3.8-Max: default max output is 16K tokens. Alibaba has not published a way to extend this significantly.
• DeepSeek V4 Pro: max output is 384K tokens — over 20 times Qwen3.8-Max's default.

If your workload involves generating long documents, code files, or extended reasoning chains that produce large outputs, DeepSeek V4 Pro has a clear advantage. For tasks where output stays under 16K tokens (most chat, Q&A, and analysis workloads), the limit does not matter.

Multimodal Capabilities

This is Qwen3.8-Max's strongest differentiator.

Qwen3.8-Max accepts text, images, and video natively. You can pass image URLs or base64-encoded images in the content array using the standard OpenAI vision format (type: "image_url"). Video input follows the same pattern. This means you can build workflows that analyze screenshots, process documents with images, or handle video frames — all through a single model call.

DeepSeek V4 Pro is text-only. No image input, no video input. If your pipeline requires vision, you need a separate model.

# Qwen3.8-Max vision request via DashScope (OpenAI-compatible)
from openai import OpenAI

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

response = client.chat.completions.create(
    model="qwen3.8-max",
    messages=[{
        "role": "user",
        "content": [
            {"type": "text", "text": "What is in this image?"},
            {"type": "image_url", "image_url": {"url": "https://example.com/photo.jpg"}}
        ]
    }],
)

For DeepSeek V4 Pro, the same SDK call works — just change the base_url and model — but you can only pass text content.

Reasoning and Thinking Mode

Both models support hybrid reasoning — the ability to toggle between fast, non-thinking responses and slower, chain-of-thought reasoning.

Qwen3.8-Max defaults to non-thinking mode. You enable thinking by setting enable_thinking: true in the extra_body parameter. When thinking is enabled, the model produces a chain-of-thought before the final answer. You can also control thinking with standard sampling parameters (temperature 0-2, top_p, penalties).

DeepSeek V4 Pro defaults to thinking mode. Every request produces chain-of-thought reasoning unless you explicitly disable it. The thinking tokens count toward your output token billing.

For routing operators, this default difference matters for cost: if your workload does not need reasoning, DeepSeek V4 Pro will burn thinking tokens unless you explicitly disable them. Qwen3.8-Max is opt-in, so non-reasoning traffic stays cheap by default.

API Access Patterns

Both models support the OpenAI chat completions format, which means any code written against the OpenAI SDK works with a base_url swap.

Qwen3.8-Max Access Routes

• DashScope OpenAI-compatible: https://dashscope.aliyuncs.com/compatible-mode/v1 with model ID qwen3.8-max
• DashScope Anthropic-compatible: available at the /apps/anthropic path
• Regional endpoints: Beijing, Singapore, Tokyo, Frankfurt, Virginia — each with workspace-scoped URLs
• No direct Qwen API: Qwen3.8-Max is only available through DashScope (Alibaba Cloud Bailian)

DeepSeek V4 Pro Access Routes

• DeepSeek direct API (OpenAI-compatible): https://api.deepseek.com with model ID deepseek-v4-pro
• DeepSeek Anthropic-compatible: https://api.deepseek.com/anthropic
• DashScope-hosted: available on DashScope with the same model ID deepseek-v4-pro
• Third-party hosts: available on OpenRouter, Together AI, and other providers

DeepSeek V4 Pro has more access routes. If provider redundancy matters to your routing setup, DeepSeek gives you more fallback options.

Benchmarks: What the Numbers Say

Independent benchmark data from Artificial Analysis (retrieved August 2026):

• Qwen3.8-Max: Intelligence Index score of 53. Top five among all tracked models, four points behind Kimi K3 (57).
• DeepSeek V4 Pro: strong performance on reasoning benchmarks. Community tests show V4 Pro excels on complex multi-step reasoning and agentic tasks.

Vendor-reported benchmarks should be treated with caution — we cite them for context, not as ground truth. In our own testing through the routing layer, both models handle complex prompts well. Qwen3.8-Max showed particular strength on Chinese-language tasks and multimodal analysis. DeepSeek V4 Pro was consistently strong on code generation and structured reasoning.

Decision Matrix: When to Pick Which

Pick Qwen3.8-Max if:

• Your workload includes image or video input (vision, document analysis, screenshot processing)
• You are primarily building for Chinese-language users
• You want thinking mode as opt-in rather than the default
• You are already on DashScope and want to stay within the Alibaba Cloud ecosystem
• You need regional endpoint options (Singapore, Tokyo, Frankfurt, Virginia)

Pick DeepSeek V4 Pro if:

• Your workload is text-only and cost is the primary concern
• You need high-volume output (up to 384K tokens per response)
• You want provider redundancy with multiple access routes (DeepSeek direct, DashScope, OpenRouter)
• Complex reasoning and code generation are your primary use cases
• You want the aggressive cache-hit pricing ($0.003625/1M input) for repeat prompts

Run both if:

• You route multimodal requests to Qwen3.8-Max and text-only requests to DeepSeek V4 Pro
• You want cost-based routing that sends high-volume output tasks to DeepSeek and short-answer tasks to either
• You need a fallback chain where if one provider is down, traffic shifts to the other

Routing Between Chinese Flagships

For operators running both models, the routing logic maps cleanly to the differences above:

1. Capability-based routing: any request with image or video content goes to Qwen3.8-Max. Text-only requests can go to either model based on cost or quality preferences.
2. Cost-based routing: for text-only workloads, DeepSeek V4 Pro at its direct API pricing is roughly 3x cheaper on output than Qwen3.8-Max on DashScope.
3. Fallback routing: if DeepSeek's API is rate-limited or unavailable, route to Qwen3.8-Max on DashScope, or vice versa. Both support the same OpenAI-compatible format, so the switch is a base_url and model change.

TheRouter routes OpenAI-compatible requests through configured providers, which means you can set up both models as provider targets and define routing rules based on content type, cost thresholds, or availability.

What We Are Watching

A few things that could shift this comparison:

• DeepSeek pricing increase: DeepSeek has announced a coming price hike. When it lands, the cost advantage over Qwen3.8-Max may narrow or disappear.
• Qwen3.8-Max open weights: Alibaba has said open weights are coming for Qwen3.8-Max, but has not published a date, license, or model card. If weights ship under Apache 2.0, self-hosting becomes an option.
• Responses API for V4 Pro: DeepSeek currently supports the Responses API only for V4 Flash, with V4 Pro support planned for August 2026. This would add another API surface for agentic workflows.

Sources

• DashScope Model Pricing — retrieved August 11, 2026
• DeepSeek API Models & Pricing — retrieved August 11, 2026
• DashScope Model Specifications — retrieved August 11, 2026
• Artificial Analysis — Qwen3.8-Max — referenced August 11, 2026
• DeepSeek V4 Preview Announcement — retrieved August 11, 2026
• Qwen3.8-Max Specs and Benchmarks (CheapestInference) — retrieved August 11, 2026