DEV Community

Cover image for What Is Qwen 3.8-Max?
Hassann
Hassann

Posted on • Originally published at apidog.com

What Is Qwen 3.8-Max?

Alibaba officially released Qwen 3.8-Max in early August 2026. It combines a 2.4T-parameter Mixture-of-Experts (MoE) architecture with 95B active parameters per token, a 1M-token context window, flat $2/$6-per-million-token pricing, and a promise to publish weights on Hugging Face and ModelScope within a week. If the weights ship as promised, this will be the first open-weight Qwen model in the Max tier.

Try Apidog today

The official announcement calls it the most capable Qwen model to date. This post focuses on the implementation details: model limits, pricing, benchmark caveats, API compatibility, and how to test the OpenAI-compatible and Anthropic-compatible endpoints. A protocol-aware client such as Apidog is useful when you need to validate both request formats with the same API key.

Qwen 3.8-Max at a glance

Spec Qwen 3.8-Max
Developer Alibaba (Qwen team)
Release Early August 2026; GA on Model Studio on August 3
Architecture MoE, built on the Qwen 3.5 foundation
Total parameters 2.4T
Active parameters 95B (~4% activation)
Context window 1,000,000 tokens
Maximum output 65,536 tokens
Modalities Text and image input; text output
Reasoning controls reasoning_effort: xhigh (default), medium, low
API model ID qwen3.8-max
Pricing $2 input / $6 output per 1M tokens
Open weights Promised for around August 10; not downloadable as of early August 2026
API protocols OpenAI-compatible and Anthropic-compatible

These details come from Alibaba’s release material and the Model Studio pricing page.

What Qwen 3.8-Max is

Qwen 3.8-Max is Alibaba’s new Qwen flagship, built on the Qwen 3.5 foundation. It replaces Qwen3.7-Max at the top of the family.

Alibaba reports substantial improvements over Qwen3.7-Max, including:

  • Terminal Bench 2.1: 74.5 → 86.6
  • PaperBench: 64.8 → 93.0

For a direct generation-to-generation comparison, see Qwen 3.8 vs Qwen 3.7 Max.

Qwen 3.8-Max model comparison

The main infrastructure detail is the MoE parameter split:

  • 2.4T total parameters
  • 95B active parameters per forward pass
  • Roughly 4% activation

That activation ratio helps explain how Alibaba can serve a 2.4T model at $2 input and $6 output per million tokens. For comparison, Kimi K3 is larger at 2.8T total parameters and 104B active parameters.

Input modalities and reasoning controls

The public API configuration lists:

  • Text input
  • Image input
  • Text output

Alibaba also demonstrates 200+ page PDF understanding and 100-hour video comprehension using “memory graphs.” Treat those as demonstrated workflows, not as additional published API input types.

Reasoning is controlled with:

{
  "reasoning_effort": "xhigh"
}
Enter fullscreen mode Exit fullscreen mode

Supported values are:

  • xhigh — default
  • medium
  • low

The API also supports enable_thinking and preserve_thinking; thinking is preserved by default.

A billing detail matters in production: thinking and non-thinking requests use the same published token rate, but thinking tokens are billed as output tokens. The xhigh default can therefore cost more than a simple input/output estimate suggests.

The first open-weight Max-class Qwen

Alibaba has released open Qwen models before, but not a Max-tier model. Qwen 3.8-Max is expected to change that: the company says weights will arrive on Hugging Face and ModelScope “next week,” approximately August 10.

Keep two constraints in mind:

  1. The weights are not available yet. As of early August 2026, this remains a release commitment, not a downloadable artifact.
  2. Self-hosting will require serious infrastructure. A 2.4T model is a multi-node deployment project even with aggressive quantization. For many teams, open weights are more likely to result in lower-cost third-party hosted options than workstation inference.

If you need access before the weights arrive, use Qwen Chat or Model Studio’s free quota. See how to use Qwen 3.8 for free.

What the benchmarks show

Alibaba published benchmark results against Claude Opus 4.8, Fable 5, GPT-5.6 Sol, and Qwen3.7-Max.

Before using these numbers for model selection, note:

  • They are vendor-run results.
  • Most coding benchmarks used the Claude Code harness for all models.
  • QwenSWEBench, QwenQoderBench, CoWorkBench, and RecreationBench are Alibaba-created benchmarks.
  • No independent results from sources such as Artificial Analysis were available at the time of writing.

Reported strengths

According to Alibaba’s table, Qwen 3.8-Max performs especially well on document, multimodal, and some agentic tasks:

  • PaperBench: 93.0 — ahead of GPT-5.6 Sol (90.5), Fable 5 (88.8), and Opus 4.8 (80.3)
  • IFBench: 82.8 — highest score in the row, ahead of Sol’s 72.7
  • Terminal Bench 2.1: 86.6 — above Opus 4.8 and Fable 5 (both 84.6), behind Sol (88.8)
  • Strong multimodal rows, including:
    • MathVision: 95.2
    • LogicVista: 91.9
    • OSWorld-Verified: 86.1
    • Leading OCR results across much of the published table

Reported weaknesses

The same table shows clear losses on core coding and high-level evaluation tasks:

  • SWE-bench Pro: 67.7
    • Fable 5: 80.0
    • Opus 4.8: 69.2
    • GPT-5.6 Sol: 64.6
  • HLE: 43.6
    • Fable 5: 53.3
    • GPT-5.6 Sol: 47.2
    • Opus 4.8: 45.7

The practical interpretation is straightforward:

  • Qwen 3.8-Max is competitive on science reasoning; it scores 92.6 on GPQA Diamond, tied with Fable 5 and just behind Sol’s 94.1.
  • It appears stronger on multimodal, OCR, document intelligence, and several agentic workflows.
  • It does not surpass Fable 5 on the published SWE-bench Pro and HLE results.

For the full table and benchmark-harness details, see the Qwen 3.8 benchmarks breakdown.

Alibaba also published showcase runs, including a 16-day autonomous coding session with 265 commits and 127 pull requests, a paper reproduction workflow, and a Tianchi contest entry. These are vendor demonstrations, not controlled independent evaluations. For coding-specific setup details, see Qwen 3.8 for coding.

How it compares with Kimi K3, Fable 5, and GPT-5.6 Sol

Qwen 3.8-Max vs. Kimi K3

Kimi K3 is the closest comparison:

Capability Qwen 3.8-Max Kimi K3
Total parameters 2.4T 2.8T
Active parameters 95B 104B
Modalities Text + image Text only
Input price $2 / 1M tokens $3 / 1M tokens
Output price $6 / 1M tokens $15 / 1M tokens
Weights Promised Already available

There is no independent direct head-to-head evaluation yet. Qwen is smaller, multimodal, and cheaper on output. K3’s immediate advantage is that its weights are already public.

See Qwen 3.8 vs Kimi K3 for a row-by-row comparison, or start with what Kimi K3 is.

Qwen 3.8-Max vs. Fable 5

Fable 5 remains ahead on Alibaba’s published core coding rows:

  • SWE-bench Pro: 80.0 vs. 67.7
  • HLE: 53.3 vs. 43.6

Qwen’s value proposition is different:

  • Lower pricing
  • 1M-token context
  • Multimodal input
  • Planned open weights

Qwen 3.8-Max vs. GPT-5.6 Sol

GPT-5.6 Sol leads on:

  • Terminal Bench: 88.8 vs. 86.6
  • GPQA: 94.1 vs. 92.6
  • HLE: 47.2 vs. 43.6

Qwen leads on:

  • PaperBench: 93.0 vs. 90.5
  • IFBench: 82.8 vs. 72.7

Qwen’s $2/$6 pricing also undercuts GPT-5.6 Terra’s $2/$12 output pricing. Sol-tier access costs more still.

For another pricing anchor, Claude Opus 5 lists at $5 input and $25 output per million tokens.

Pricing: $2/$6 across the full 1M-token context

Qwen 3.8-Max pricing is:

  • $2 per million input tokens
  • $6 per million output tokens
  • Flat from token zero through the 1M-token context limit

The flat context tier is unusual. Many long-context APIs increase token prices after a prompt crosses a threshold. Qwen 3.8-Max does not.

It is also cheaper than Qwen3.7-Max’s standard list price of $2.5/$7.5 per million tokens. Qwen3.7-Max currently has a 50% promotion at $1.25/$3.75, however, so it remains relevant for cost-sensitive workloads.

Cache pricing

For repeated-prefix workloads:

  • Cache hits cost 10% of the input rate
  • Explicit cache creation costs 125% of the input rate

Model Studio also provides a free quota:

  • 1M tokens
  • Singapore region only
  • Valid for 90 days

For per-task pricing examples and reasoning-token cost considerations, see the Qwen 3.8 pricing guide.

How to access Qwen 3.8-Max

There are five primary access routes.

1. Use Qwen Chat

Qwen Chat is the fastest route for interactive testing and does not require an API key.

2. Use the Model Studio API

Get an API key from home.qwencloud.com, export it, and call the qwen3.8-max model.

export DASHSCOPE_API_KEY="your-api-key"
Enter fullscreen mode Exit fullscreen mode

Available OpenAI-compatible regional base URLs:

Beijing:
https://dashscope.aliyuncs.com/compatible-mode/v1

Singapore:
https://dashscope-intl.aliyuncs.com/compatible-mode/v1

US Virginia:
https://dashscope-us.aliyuncs.com/compatible-mode/v1
Enter fullscreen mode Exit fullscreen mode

The Model Studio models page lists Qwen 3.8-Max in its recommended model stack.

OpenAI-compatible request example 

curl https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions \
  -H "Authorization: Bearer $DASHSCOPE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-max",
    "messages": [
      {
        "role": "user",
        "content": "Explain the trade-offs of a Mixture-of-Experts model."
      }
    ],
    "reasoning_effort": "medium"
  }'
Enter fullscreen mode Exit fullscreen mode

3. Use the Anthropic-compatible endpoint

The Anthropic-compatible endpoint is:

https://dashscope-intl.aliyuncs.com/apps/anthropic
Enter fullscreen mode Exit fullscreen mode

It implements the Anthropic Messages protocol. This allows Claude-targeted tooling to point at Qwen with environment configuration rather than a full client rewrite.

4. Use coding harnesses

Alibaba published configurations for:

  • Claude Code
  • Codex
  • Qoder
  • Qwen Code
  • OpenClaw

For Claude Code, configure the Anthropic-compatible endpoint and model ID:

export ANTHROPIC_BASE_URL="https://dashscope-intl.aliyuncs.com/apps/anthropic"
export ANTHROPIC_MODEL="qwen3.8-max"
Enter fullscreen mode Exit fullscreen mode

Alibaba also used the Claude Code harness for many of its coding benchmarks, making this setup relevant if you want to reproduce comparable workflows.

5. Wait for open weights

Alibaba says Qwen 3.8-Max weights will be published on Hugging Face and ModelScope around August 10. They were not downloadable at the time of writing.

For copy-paste Python and cURL setup across the API routes, see the Qwen 3.8 API guide.

Test both API protocols

Because Qwen 3.8-Max supports both OpenAI-compatible and Anthropic-compatible APIs, test protocol behavior before migrating a production client.

A useful validation workflow is:

  1. Save each regional DashScope URL as an environment.
  2. Send the same prompt to the OpenAI-compatible endpoint.
  3. Send an equivalent request to the Anthropic-compatible endpoint.
  4. Compare output content, tool behavior, streaming events, token usage, and latency.
  5. Test each reasoning_effort level separately.

In Apidog, you can store regional URLs as environments, switch regions without editing individual requests, and compare the two protocol formats side by side. You can also inspect SSE streaming responses to observe reasoning_content deltas before the final answer.

That is especially useful for estimating how much output is generated by the default xhigh reasoning mode.

Download Apidog to test qwen3.8-max against qwen3.7-max or kimi-k3 using identical prompts and request settings.

Where Qwen 3.8-Max fits in the lineup

Qwen 3.8-Max sits above Qwen3.7-Max and the plus-tier Qwen models.

A practical selection rule:

  • Choose Qwen 3.8-Max for multimodal tasks, document intelligence, and high-end agentic workloads.
  • Choose Qwen3.7-Max when its 50% promotion provides the better value.
  • Choose plus-tier models for routine, lower-cost workloads.

For a broader comparison, see the best Qwen models guide.

FAQ

Is Qwen 3.8-Max open source?

Not yet. Alibaba promised weights for Hugging Face and ModelScope “next week” from its early August 2026 launch. Until the files are published, use the API or Qwen Chat. For no-cost options, see how to use Qwen 3.8 for free.

How much does Qwen 3.8-Max cost?

It costs $2 per million input tokens and $6 per million output tokens across the full 1M-token context window.

Cache hits cost 10% of the input rate. Thinking tokens count as output tokens, so reasoning-heavy requests can cost more than a basic prompt-token estimate.

Is Qwen 3.8-Max better than Kimi K3?

There is no independent head-to-head evaluation yet. On paper, Qwen 3.8-Max is smaller, supports image input, and costs less on output. Kimi K3’s main advantage is that its weights are already public.

Can I use Qwen 3.8-Max in Claude Code?

Yes. Set ANTHROPIC_BASE_URL to DashScope’s Anthropic-compatible endpoint and set ANTHROPIC_MODEL=qwen3.8-max. Alibaba also published configurations for Codex, Qoder, Qwen Code, and OpenClaw.

What to do next

Qwen 3.8-Max is available as a general API release across three regions, with published pricing and a detailed—though vendor-run—benchmark table. Its strongest reported areas are multimodal, document, OCR, and selected agentic workflows. It does not outperform Fable 5 on the published core coding benchmarks, but it offers a lower $2/$6 price point and a planned open-weight release.

The best evaluation method is to test your own workload:

  1. Claim the free quota.
  2. Configure both protocol endpoints.
  3. Run your production prompts through each.
  4. Compare quality, latency, reasoning-token usage, and cost.
  5. Revisit deployment options when the weights are published.

Start with the Qwen 3.8 API setup guide, then check whether the weights arrived on schedule around August 10.

Top comments (0)