Alibaba

Qwen3.7-Max

Provider

Alibaba

Status

available

Context

1,000,000 tok

SWE-bench

80.4%

Price

$2.5 / $7.5 /MTok

Qwen3.7-Max is Alibaba’s flagship Qwen model, announced on 20 May 2026 at the Alibaba Cloud Summit, with the API going live a day earlier. It is a reasoning-native agent model with a 1-million-token context window (MarkTechPost) — and, notably, the first Qwen flagship to ship closed-weight and API-only, a sharp break from Alibaba’s open-weight tradition.

It is the strongest Qwen yet on benchmarks: 80.4% on SWE-bench Verified on Alibaba’s own testing, and at launch the highest-placed Chinese model on Artificial Analysis’s intelligence index (Artificial Analysis). That puts it within striking distance of the US frontier (GPT-5.5, Claude Opus 4.8, Gemini 3.x) at roughly half their price, while the open Qwen 3.6 models remain the self-hostable alternative.

Update, 11 September 2026. Artificial Analysis re-based its Intelligence Index on 7 September (v4.3: GPQA Diamond out, Terminal-Bench 4.0 and AutomationBench-AA in), pulling every model’s score down. Qwen3.7-Max now reads 30, 68th of 200, at $1.15 per task. Index figures elsewhere on this page are the readings at the time they were written, on the old scale, and are kept as dated.

Quick specs

VIEW QWEN3.7-MAX →

What Qwen3.7-Max is

Qwen3.7-Max is positioned as a reasoning agent model: extended thinking is on by default, and it is built for long-horizon, tool-heavy work (MarkTechPost). Its 1-million-token context window is a fourfold jump over the 256K on the preceding Qwen3.6-Max-Preview, and it supports function calling, MCP integration and external agent harnesses — Alibaba demonstrated it running autonomously for extended sessions on a single task (VentureBeat).

Alibaba does not disclose the parameter count for its Max tier, and whether 3.7-Max is dense or mixture-of-experts is unconfirmed. The Max line is text-only; multimodal and vision capability live in the sibling Qwen3.7-Plus-Preview released alongside it.

The headline strategic fact is the licence. Where Alibaba built its reputation on Apache-2.0 open weights, it has now moved the flagship behind a closed, API-only wall — a two-track strategy of an open mid-tier (Qwen 3.6) and a proprietary frontier, and a shift that drew criticism from the open-source community.

Benchmark performance

Alibaba published a broad vendor benchmark table; independent leaderboards temper it.

The pattern is a frontier-adjacent agentic and coding model that leads or ties on several vendor benchmarks while trailing GPT-5.5, Claude Opus 4.8 and Gemini 3.x on the independent intelligence index. Two caveats from Artificial Analysis: it is verbose — generating roughly four times the median output tokens in evaluation, which erodes its price advantage — and its raw factual accuracy fell versus the predecessor, with lower hallucination coming largely from higher abstention. See best AI models for cross-model standings.

Pricing and access

Qwen3.7-Max is priced at $2.50 input / $7.50 output per million tokens on Alibaba Cloud Model Studio, with cached input at $0.25 (Artificial Analysis). OpenRouter listed a launch promotion of $1.25 / $3.75 — a 50% discount on the same base rate, not a separate tier. The API model ID is qwen3.7-max, with OpenAI- and Anthropic-compatible endpoints; it is also available via Together AI and in the consumer Qwen app.

There are no open weights — this is a closed model, and hosted access routes through Alibaba Cloud’s Singapore international region (China jurisdiction). For self-hostable Qwen, see the Apache-2.0 Qwen 3.6 models.

How Qwen3.7-Max compares

• vs the proprietary frontier — On the independent intelligence index it sits just below GPT-5.5, Claude Opus and Gemini 3.x, but leads or ties them on several agentic, coding and STEM vendor benchmarks, at roughly half the price.
• vs other Chinese models — At launch it was the top-placed Chinese model on Artificial Analysis, competing with DeepSeek V4, MiniMax M3 and Kimi — though DeepSeek undercuts it on price and ships open weights.
• vs the open Qwen 3.6 — 3.7-Max is more capable and has a 1M context, but is closed and API-only; the open 3.6 models trade some capability for Apache-2.0 self-hosting.

Known limitations

Closed weights, API-only — no self-hosting, a departure from Alibaba’s open tradition. China-hosted — hosted access carries data-residency, compliance and content-control considerations. Verbose — high output-token counts raise real-world cost despite the modest headline price. Undisclosed architecture — no published parameter count or AIME figure, and the independent intelligence-index figure shifted between launch and the current revised index. Text-only — vision lives in the separate Qwen3.7-Plus-Preview.

FAQ

What is Qwen3.7-Max?

Qwen3.7-Max is Alibaba’s flagship Qwen model, announced 20 May 2026 — a closed-weight, API-only reasoning agent with a 1-million-token context window, built for long-horizon agentic coding and tool use.

Is Qwen3.7-Max open source?

No. It is the first Qwen flagship to ship closed-weight and API-only, available through Alibaba Cloud Model Studio rather than as downloadable weights. For open Qwen, see the Apache-2.0 Qwen 3.6 models.

How much does Qwen3.7-Max cost?

$2.50 per million input tokens and $7.50 per million output on Alibaba Cloud, with cached input at $0.25. OpenRouter listed a 50%-off launch promotion of $1.25 / $3.75. Note that the model is verbose, which raises real-world cost.

How good is Qwen3.7-Max?

Very strong — 80.4% SWE-bench Verified (vendor) and the highest-placed Chinese model on Artificial Analysis at launch — but it trails GPT-5.5, Claude Opus 4.8 and Gemini 3.x on the independent intelligence index.

What is the difference between Qwen3.7-Max and Qwen 3.6?

Qwen 3.6 is Alibaba’s open-weight generation (Apache 2.0, self-hostable). Qwen3.7-Max is the closed, API-only flagship — more capable, with a 1M-token context, but not downloadable.

Last verified 19 June 2026. The 20 May 2026 launch, closed-weight/API-only status, 1M context and pricing are confirmed by Alibaba’s materials, Artificial Analysis and OpenRouter. Benchmark figures are largely vendor-reported (Alibaba’s own testing); the Artificial Analysis intelligence-index score shifted from 56.6 (launch, v4.0) to 46 (current, revised index) and is shown with both. Confirm figures against current leaderboards before relying on them.