AI/TLDR

Qwen3.7-Max

Alibaba's flagship agent model with a 1M-token context and 35-hour autonomy

Overview

Qwen3.7-Max is the current flagship of Alibaba's Qwen-Max line, announced by the Qwen team around the 2026 Alibaba Cloud Summit (May 19-21, 2026) and made available the same week through Alibaba Cloud Model Studio. It is a proprietary, closed-weight model positioned explicitly for the "agent era": Alibaba describes it as its most capable agent foundation to date, equally suited to writing and debugging code, automating office workflows, and sustaining long-horizon tasks that span hundreds or thousands of steps.

Technically, Qwen3.7-Max ships with a 1M-token context window (double the 256K of the earlier Qwen3.6 Max Preview) and a maximum output of 65,536 tokens. It is a text-only model — there is no image, audio, or video input — and uses native extended-thinking (chain-of-thought) reasoning. Its headline demonstration is a 35-hour, fully autonomous kernel-optimization run comprising over 1,000 tool calls, which Alibaba cites as evidence that the model holds coherent state across extremely long agentic horizons. It supports OpenAI- and Anthropic-compatible APIs and can be driven by external harnesses such as Claude Code.

On the Artificial Analysis Intelligence Index v4.0 it scored 56.6 at launch — ranked #5 overall and the highest-placed Chinese model at the time — a 4.8-point gain over Qwen3.6 Max Preview. It posts strong coding-agent results (80.4 on SWE-Bench Verified, 69.7 on Terminal-Bench 2.0, 76.4 on MCP-Atlas) and 92.4 on GPQA Diamond. Pricing is $2.50 / $7.50 per million input/output tokens (with $0.25 cached input), and Alibaba ran a 50% launch promotion at $1.25 / $3.75. The model is also cross-listed on OpenRouter and Together AI.

Released2026-05
LicenseProprietary (closed weights); API-only via Alibaba Cloud Model Studio
WeightsAPI only
ParametersNot disclosed
Context1M
Max output65,536 tokens
ArchitectureProprietary reasoning ("extended thinking") model. Alibaba has not published a parameter count or full architecture; coverage places it within Qwen's Mixture-of-Experts (MoE) family, but this is not officially confirmed.
Knowledge cutoffNot disclosed
ModalitiesText
StatusAvailable

Benchmarks

  1. Artificial Analysis Intelligence Index v4.056.6index
  2. SWE-Bench Verified80.4%
  3. SWE-Bench Pro60.6%
  4. Terminal-Bench 2.0 (Terminus)69.7%
  5. MCP-Atlas76.4%
  6. GPQA Diamond92.4%
  7. Humanity's Last Exam41.4%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

Pricing

Input$2.50 / 1M tokens per 1M tokens
Cached input$0.25 / 1M tokens per 1M tokens
Output$7.50 / 1M tokens per 1M tokens

List price. Alibaba ran a 50% launch promotion at $1.25 input / $3.75 output per 1M tokens, also reflected on OpenRouter and Together AI.

Pricing source ↗

Strengths

  • Very long context: 1M-token window for whole-repo and long-document agent tasks
  • Long-horizon autonomy: cited 35-hour autonomous run with 1,000+ tool calls without losing state
  • Strong coding-agent benchmarks (SWE-Bench Verified 80.4, Terminal-Bench 2.0 69.7, MCP-Atlas 76.4)
  • Top-tier reasoning: 92.4 on GPQA Diamond; #5 / highest Chinese model on the AA Intelligence Index at launch
  • Aggressive pricing for a frontier model, with a 90% cached-input discount that suits repeated agent context
  • OpenAI- and Anthropic-compatible APIs; works with external harnesses like Claude Code

Best for

  • Autonomous coding agents: multi-file refactors, debugging, and long-running build/test loops
  • Long-horizon agentic workflows that chain hundreds-to-thousands of tool calls
  • Whole-repository and large-document analysis enabled by the 1M-token context
  • Office and productivity automation (spreadsheets, multi-step task pipelines)
  • Scientific and technical reasoning, including STEM-heavy QA and math
  • Cost-sensitive production deployments via cached-input pricing on Alibaba Cloud, OpenRouter, or Together AI

How to access

ProviderModel ID
Alibaba Cloud Model Studio ↗qwen3.7-max
OpenRouter ↗qwen/qwen3.7-max
Together AI ↗Qwen/Qwen3.7-Max

Qwen-Max — every version

The full lineage of the Qwen-Max line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.

VersionReleasedContextLicense
Qwen3.7-Maxcurrent2026-05Proprietary
Qwen3-Max2025-09Proprietary
Qwen2.5-Max2025-01-29Proprietary

FAQ

Is Qwen3.7-Max open source or open weights?

No. Qwen3.7-Max is a proprietary, closed-weight model. As of its May 2026 launch there were no Qwen3.7-Max weights on Hugging Face; it is available only through APIs such as Alibaba Cloud Model Studio, OpenRouter, and Together AI.

What is Qwen3.7-Max's context window and maximum output?

It has a 1M-token context window — double the 256K of the earlier Qwen3.6 Max Preview — and a maximum output of 65,536 tokens per request.

Does Qwen3.7-Max support images, audio, or video?

No. Qwen3.7-Max is text-only for both input and output. There is no native vision, audio, or video support in this model.

How much does Qwen3.7-Max cost?

List pricing is $2.50 per million input tokens and $7.50 per million output tokens, with cached input at $0.25 per million. At launch Alibaba offered a 50% promotion at $1.25 input / $3.75 output per million tokens, also reflected on OpenRouter and Together AI.