Qwen3.7-Max: Alibaba's Agent-First 1M-Context Flagship

Alibaba announced Qwen3.7-Max on May 20, 2026 at its Cloud Summit, with the commercial API going live on DashScope a day earlier. The framing is the interesting part here: rather than a chatbot that happens to reason, Alibaba positions it as an agent-first flagship, built for long-running, multi-step workflows instead of single-prompt question answering. It ships with a 1M-token context window and a native extended-thinking mode.

Update (July 2026): Alibaba has since moved its flagship forward. It previewed Qwen3.8-Max on July 19, 2026, a 2.4-trillion-parameter multimodal model it bills as second only to Fable 5. The coverage below reflects Qwen3.7-Max.

An Agent-first Flagship

The “Max” tier has always been Alibaba’s largest, most capable line, and 3.7 leans hard into the agentic direction the whole field has moved toward. Extended thinking is built in rather than bolted on, and the 1M context is reworked so retrieval stays useful at the tail of the window instead of degrading into noise, which matters a lot when an agent is dozens of tool calls deep and the relevant detail is buried 700K tokens back.

The Benchmarks

The numbers Alibaba leads with are agentic and competition-grade. On SWE-Pro it scores 60.6, the top of its comparison table, ahead of DeepSeek V4-Pro and Moonshot’s K2.6 Thinking. On Terminal-Bench 2.0 it posts 69.7, again ahead of DeepSeek V4-Pro (67.9) and Claude Opus 4.6 (65.4). It hits 92.4 on GPQA Diamond, 41.4 on Humanity’s Last Exam (ahead of Opus 4.6’s 40.0), and a category-leading 97.1 on HMMT February 2026 competition math.

Standard caveat applies: these are vendor-selected comparisons, so treat them as a strong starting point and benchmark on your own agent traces. But the consistent story across them is that Qwen3.7-Max is competitive with, and on several agentic tasks ahead of, the current Western frontier.

What Developers Are Saying

The early developer reception has been unusually warm, and it centers on endurance rather than raw benchmark scores. Reviewers testing it on agentic coding have called the experience genuinely impressive, with the standout being how long it holds a thread without losing the plot, the exact failure mode that makes most agents frustrating to run unattended.

The headline anecdote comes from Alibaba’s own internal testing: the model reportedly ran autonomously for roughly 35 hours, firing 1,158 tool calls and 432 kernel evaluations to optimize a GPU attention kernel it had never seen, landing a 10× speedup over the reference. Worth being honest about what that is: a vendor benchmark, on Alibaba’s hardware, not yet independently reproduced. Treat it as a strong directional signal that Qwen is optimizing for long-horizon reliability, not a guarantee your agent runs for 35 hours on day one.

There are two practical things developers keep flagging in their writeups. First, it natively speaks the Anthropic API protocol, so dropping it into a Claude Code or Anthropic-SDK setup is close to a no-op, a real reduction in switching cost. Second, hands-on reviews note it holds performance across harnesses (Claude Code, Qwen Code, custom frameworks) and outputs around 192 tokens per second, quick for a reasoning model, which matters when an agent is making hundreds of decisions.

Closed Weights, API-only

Worth being clear, because it trips people up: Qwen3.7-Max is closed-weight and API-only, so you can’t download it and run it yourself. That sets it apart from Alibaba’s genuinely open Qwen line, the smaller dense and MoE releases you can self-host and fine-tune. If open weights are a hard requirement, the Max tier isn’t your model; one of the open Qwen releases or DeepSeek V4 is, and our roundup of the best open source models for coding ranks them on coding work specifically.

Pricing

On DashScope it runs $2.50 per million input tokens and $7.50 per million output, with cached input at roughly $0.25, a 90% discount that’s genuinely useful for long-context agentic loops, where you’re replaying a large, mostly-static context across many turns. That puts it above the cheapest open-weight options but competitive with the proprietary frontier, and the cache economics specifically reward the long-horizon workloads it’s built for. It was cross-listed on OpenRouter and Together AI from day one.

Where It Fits

Qwen3.7-Max is strongest where the work is agentic and multi-step (tool-calling pipelines, terminal and coding agents, long-horizon tasks), and it keeps Qwen’s longstanding edge on Chinese-language reasoning. For developers outside China, the one thing I’d still test before committing is international API reliability; Alibaba’s global infrastructure has improved a lot but historically trailed OpenAI and Anthropic on uptime consistency.

The broader signal is the same one the rest of 2026 keeps sending: the agentic-coding frontier is crowded now, with Alibaba, DeepSeek, Anthropic, Google, and OpenAI all landing within a few benchmark points of each other. Good news if you’re building: more leverage, lower prices, and a capability floor that keeps rising.