Qwen3.7-Max

Overview

Qwen3.7-Max is Alibaba Cloud's flagship large language model, unveiled at the Alibaba Cloud Summit in Hangzhou and served from Alibaba Cloud Model Studio as qwen3.7-max, with a dated snapshot qwen3.7-max-2026-05-20 and a subsequent qwen3.7-max-2026-06-08.

Two Alibaba datelines bracket the launch. The company's summit announcement is bylined May 20, 2026 and reports that "Alibaba on Wednesday launched" the model — May 20, 2026 was indeed a Wednesday. The technical launch post, "Qwen3.7: The Agent Frontier", is bylined May 21, 2026. This page dates the model to May 20, the date on the announcement and on the -2026-05-20 snapshot ID.

The most consequential fact about it is not a benchmark. Qwen3.7-Max is proprietary and closed-weight. There is no repository for it on github.com/QwenLM and no weights on huggingface.co/Qwen. For a series whose reputation was built on shipping frontier-adjacent weights under Apache 2.0, this is a real shift in posture: the flagship is now an API product, and only the sub-flagship Qwen3.6 line remains openly licensed.

Alibaba positions the model as agent-first. Its documentation places it at the top of the "recommended models" table with a 1M token context window, and its deep-thinking guide lists it as a hybrid thinking model with thinking enabled by default — reasoning happens unless you explicitly disable it.

Capabilities

• Hybrid thinking, on by default: qwen3.7-max reasons before answering unless enable_thinking is set to false. A thinking-only variant (qwen3.7-max-preview, qwen3.7-max-2026-05-17) exists for workloads that always want deliberation.
• 1M token context: Alibaba's own documentation lists a 1M context window for both qwen3.7-max and qwen3.7-plus, framing it as "roughly 750,000 English words, or 8-10 novels."
• Text-only API surface: Alibaba's deep thinking FAQ asks "Can I upload images or documents as input?" and answers: "These models accept text only. Qwen3-VL and QVQ support deep thinking on images." Vision work routes to Qwen3-VL and QVQ.
• Context caching: Model Studio tags a context caching discount on qwen3.7-max, qwen3.7-max-2026-05-20, and qwen3.7-max-2026-06-08. The thinking-only qwen3.7-max-preview and qwen3.7-max-2026-05-17 snapshots carry no such tag on the pricing page.
• Dated snapshots: Pinning to qwen3.7-max-2026-05-20 or qwen3.7-max-2026-06-08 gives reproducible behavior across deployments.
• Regional deployments: Separate International (Singapore) and Chinese Mainland (Beijing) endpoints with distinct rate cards.

Technical Specifications

• Model IDs: qwen3.7-max (rolling alias), qwen3.7-max-2026-05-20, qwen3.7-max-2026-06-08
• Thinking-only IDs: qwen3.7-max-preview, qwen3.7-max-2026-05-17
• Context window: 1M tokens
• Thinking: hybrid mode, enabled by default; disable via enable_thinking: false
• Modalities: text only
• Weights: not published — closed, API-only
• License: proprietary
• Platforms: Alibaba Cloud Model Studio and DashScope

Alibaba has not published a parameter count, architecture description, training-data size, or knowledge cutoff for Qwen3.7-Max on any primary source. Those fields are deliberately absent from this page rather than filled with estimates.

The open-weight line: Qwen3.6

The Qwen3.6 generation continues under Apache 2.0, and is where self-hosting teams should look:

The QwenLM repository states flatly: "All our open-weight models are licensed under Apache 2.0." A hosted qwen3.6-flash is also available on Model Studio.

qwen3.7-plus

A lower-cost sibling in the same generation — qwen3.7-plus is a rolling alias for qwen3.7-plus-2026-05-26 — also listed at a 1M token context window. Both its input and its output pricing are tiered by the number of input tokens in the request, rather than flat. Crossing the 256K boundary triples the input rate and triples the output rate together, so a long-context agent loop costs considerably more per call than the headline tier suggests.

Use Cases

• Long-horizon agents: Hybrid thinking on by default plus a million tokens of context suits agents that plan, call tools, and revise across many steps.
• Whole-codebase reasoning: A 1M window holds a substantial repository in a single request, removing much of the retrieval scaffolding.
• Document-scale analysis: Contracts, filings, and research corpora processed without chunking.
• Multilingual enterprise workloads: The Qwen series' long-standing strength in Chinese and English, delivered as a managed API.
• Self-hosted deployments — via Qwen3.6: If open weights are a hard requirement, Qwen3.6-27B and Qwen3.6-35B-A3B are the models to build on. Qwen3.7-Max cannot be self-hosted.

Performance / Benchmarks

Alibaba published a first-party benchmark table with the launch post, "Qwen3.7: The Agent Frontier" (Alibaba Cloud Community, May 21, 2026), which opens "Today we introduce Qwen3.7-Max, our latest proprietary model designed for the agent era."

The post is easy to miss: the qwen.ai blog renders entirely client-side and exposes no static text to fetch, which is why the article is often reported as nonexistent. The same article is served statically on alibabacloud.com. The scores below are transcribed from it. All figures are Alibaba-reported vendor numbers, measured across a mix of public and Alibaba-internal benchmarks.

Where Alibaba cited a competitor's score alongside its own, that comparator is reproduced in the third column exactly as the post states it.

Coding agents

General-purpose agents

Reasoning

General capabilities and multilingualism

Three caveats travel with these numbers. First, they are vendor-reported and unreplicated: treat them as an upper bound and validate on your own task set. Second, several benchmarks Alibaba cites in the same post — QwenSVG, QwenClawBench, CoWorkBench, QwenWebDev, and QwenWorldBench — are Alibaba-internal constructions, so a strong showing on them is not independently meaningful. Third, the harnesses matter: Alibaba's own footnotes specify a Harbor/Terminus-2 harness with a five-hour timeout for Terminal-Bench 2.0, an internal bash-plus-file-edit scaffold for the SWE-Bench series, and a gemini-2.5-pro judge for MCP-Atlas. Alibaba also notes that on SWE-bench Pro it "corrected" problematic tasks and re-evaluated all baselines on the refined benchmark.

Alibaba frames the breadth as the point: the scores are "drawn from a wide variety of agent scaffolds," and the model is claimed to perform consistently "across Claude Code, OpenClaw, Qwen Code, and custom tool-use frameworks."

For open-weight comparisons, see DeepSeek V4, Kimi K2.6, and GLM-5.2, all of which publish model cards with evaluation tables and downloadable weights.

Limitations

• Closed weights: No self-hosting, no fine-tuning on your own infrastructure, no air-gapped deployment. This is the single biggest change from prior Qwen flagships.
• Text only: The Max snapshots support only the text API. Vision requires a different model.
• No published architecture: Parameter count, active parameters, expert configuration, and training-token count are all unstated.
• No published knowledge cutoff: Ground time-sensitive queries with retrieval or tool use.
• Benchmarks are vendor-reported and partly internal: Alibaba publishes an evaluation table, but it is first-party, unreplicated, and leans on several Alibaba-built benchmarks (QwenSVG, QwenClawBench, CoWorkBench, QwenWorldBench) that no third party can audit.
• Regional price divergence, and two currencies: The Singapore endpoint is quoted natively in USD, the Beijing endpoint natively in CNY. Cost models must account for which region serves the request and for the exchange rate Alibaba applies when it restates one in the other.
• Tiered pricing on qwen3.7-plus: Crossing 256K input tokens triples both the input and the output rate. qwen3.7-max is flat across its full 1M window, so the cheaper model is not uniformly cheaper.
• Thinking is on by default: Latency and output-token spend will be higher than a non-reasoning baseline unless enable_thinking is explicitly disabled.

Pricing & Access

Alibaba quotes each region in its own currency, and the two rate cards are not conversions of one another. The International (Singapore) card is natively in US dollars. The Chinese mainland (North China 2, Beijing) card is natively in Chinese yuan. Each of Alibaba's two pricing pages restates the other region's figures in its own currency — which is why the English page shows the mainland at "$1.65 / $4.951" and the Chinese page shows the US endpoint at "18.736 元 / 56.207 元". The round numbers are the native quotes; the long decimals are conversions. Prices below are per 1M tokens, as the vendor states them.

Both qwen3.7-max and qwen3.7-plus support a context caching discount. qwen3.7-max is priced flat across its entire 1M window; qwen3.7-plus is tiered on both input and output.

International (Singapore) — quoted in USD

Source: Alibaba Cloud Model Studio pricing.

A US (Virginia) endpoint, qwen3.7-max-us, is listed at the same $2.50 / $7.50.

Chinese mainland — North China 2 (Beijing) — quoted in CNY

Source: 阿里云百炼模型计费 ("Alibaba Cloud Model Studio model pricing"), section 华北2（北京）("North China 2 (Beijing)").

The Chinese page renders the qwen3.7-max row as 12 元 and 36 元 under the column headers 输入单价（每百万 Token）and 输出单价（每百万 Token）— "input unit price (per million tokens)" and "output unit price (per million tokens)". Alibaba's English page restates that same row as $1.65 / $4.951, and the qwen3.7-plus tiers as $0.276 / $1.101 and $0.826 / $3.301. Those USD figures are conversions, not the vendor's native quote.

Access options:

• Alibaba Cloud Model Studio — the primary managed API
• DashScope — Alibaba's model-serving platform
• Third-party aggregators — resell the hosted API; there is no self-hosted path
• Open weights — only for the Qwen3.6 line, on Hugging Face and ModelScope under Apache 2.0

Ecosystem & Tools

• Qwen3.7: The Agent Frontier — Alibaba's launch post and the only primary source for its benchmark table
• Alibaba Cloud Model Studio — supported models, capabilities, and snapshot IDs
• Model Studio pricing (English) — per-region rate cards, USD
• 百炼模型计费 (Chinese) — the mainland rate card as Alibaba natively quotes it, in CNY
• Deep thinking guide — hybrid thinking configuration and the enable_thinking parameter
• QwenLM/Qwen3.6 on GitHub — the open-weight generation
• Qwen on Hugging Face — Apache 2.0 weights for Qwen3.6 and earlier
• ModelScope — Alibaba's model community

Community & Resources

• Supported models and capabilities overview
• Text generation models on Model Studio
• Qwen3.6 repository
• Qwen
• Compare with DeepSeek V4, Kimi K2.6, GLM-5.2, and Tencent Hy3