Qwen3.7-Max Just Hit #1 on Hacker News — Here’s Why Developers Are Paying Attention

Qwen3.7-Max dropped yesterday. It’s not on OpenRouter. It’s not on your favorite AI dashboard. But it’s already beating Opus on coding benchmarks — and it’s from the Chinese team most developers keep ignoring.

Alibaba’s Qwen team released Qwen3.7-Max on May 20, 2026. Within hours, it hit #1 on Hacker News with 573 points and 223 comments. The developer community didn’t just upvote it — they started benchmarking it against their own tools.

The reason: this model ran a 35-hour autonomous kernel optimization task. Over 1,000 tool calls. Zero human intervention. It achieved a 10x speedup over the reference implementation.

I went through the full blog post, the HN thread, and the benchmark data. Here’s what actually matters.

TL;DR

• Qwen3.7-Max released May 20, 2026 and immediately hit #1 on HN (573 pts, 223 comments)
• Beats or matches Opus-4.6 Max on most coding and reasoning benchmarks
• Ran a 35-hour autonomous kernel optimization with 1,158 tool calls, achieving 10x speedup
• Works with Claude Code via Anthropic API compatibility
• Not open-weight — cloud-only via Alibaba Cloud Model Studio
• Not on OpenRouter yet — signup friction is real
• Best for: developers who want frontier agent capabilities without being locked into US providers

The Benchmark Table Everyone’s Screenshotting

Forget the marketing. Here’s the data:

| Benchmark          | Opus-4.6 Max | DS-V4-Pro Max | Qwen3.6-Plus | Qwen3.7-Max |
|--------------------|--------------|---------------|--------------|-------------|
| Terminal Bench 2.0 | 65.4         | 67.9          | 61.6         | 69.7        |
| SWE-Verified       | 80.8         | 80.6          | 78.8         | 80.4        |
| SWE-Pro            | 57.3         | 59.0          | 56.6         | 60.6        |
| SWE-Multilingual   | 77.5         | 76.2          | 73.8         | 78.3        |
| SciCode            | 51.9         | --            | 41.4         | 53.5        |
| GPQA Diamond       | 91.3         | 90.1          | 90.4         | 92.4        |
| HLE                | 40.0         | 37.7          | 28.8         | 41.4        |
| HMMT 2026 Feb      | 96.2         | 95.2          | 87.8         | 97.1        |
| IMOAnswerBench     | 75.3         | 89.8          | 83.8         | 90.0        |
| Kernel Bench L3    | 2.63x/98%    | 1.07x/54%     | 1.03x/48%    | 1.98x/96%   |

On HMMT 2026 Feb — a brutal math competition — it scores 97.1. Opus-4.6 gets 96.2. DeepSeek V4 Pro gets 95.2.

On SWE-Pro, it hits 60.6. That’s ahead of every model they tested.

On Terminal Bench 2.0, it scores 69.7 — beating DeepSeek V4 Pro Max (67.9) and everyone else.

But the benchmarks aren’t the most interesting part.

The 35-Hour Kernel Optimization Run

This is what made the HN thread explode.

Alibaba gave Qwen3.7-Max a task: optimize the Extend Attention kernel in SGLang. It’s a memory-bound, latency-critical operator for LLM serving. The model had never seen the hardware platform before — T-Head ZW-M90 PPUs, a chip architecture that didn’t exist in its training data. No profiling data. No example kernels. Just a task description and an evaluation script.

What happened:

• 35 hours of continuous autonomous execution
• 1,158 tool calls
• 432 kernel evaluations
• Final result: 10x geometric mean speedup over the Triton reference

The model was still finding real improvements after 30+ hours. It diagnosed compilation failures, fixed correctness bugs, profiled bottlenecks, and redesigned the kernel architecture. All on its own.

They ran the same task with other models under identical conditions:

| Model           | Speedup | Notes            |
|-----------------|---------|------------------|
| Qwen3.7-Max     | 10.0x   | Full 35-hour run |
| GLM 5.1         | 7.3x    | Stopped early    |
| Kimi K2.6       | 5.0x    | Stopped early    |
| DeepSeek V4 Pro | 3.3x    | Stopped early    |
| Qwen3.6-Plus    | 1.1x    | Barely improved  |

The other models stopped because the agent decided it couldn’t make more progress. Qwen3.7-Max kept going.

On KernelBench L3 — a broader test across 50 GPU kernel problems — Qwen3.7-Max produced accelerated kernels for 96% of scenarios. Opus-4.6 managed 98%. GLM 5.1 got 78%. DeepSeek V4 Pro got 54%.

It Works With Your Existing Setup

Qwen3.7-Max supports the Anthropic API protocol. That means you can plug it into Claude Code right now:

npm install -g @anthropic-ai/claude-code

export ANTHROPIC_MODEL="qwen3.7-max"
export ANTHROPIC_SMALL_FAST_MODEL="qwen3.7-max"
export ANTHROPIC_BASE_URL=https://dashscope-intl.alibabacloud.com/apps/anthropic
export ANTHROPIC_AUTH_TOKEN=<your_api_key>

claude

It also works with OpenClaw, Qwen Code (their open-source terminal agent — 24K GitHub stars), and any framework that supports OpenAI-compatible APIs.

One HN commenter was already running Qwen3.6 through llama.cpp with OpenCode as a free Claude Code alternative. Qwen3.7-Max is a significant step up.

The Catch

Qwen3.7-Max is not open-weight. It’s proprietary, cloud-only, available exclusively through Alibaba Cloud Model Studio. You can’t download it. You can’t self-host it.

It’s not on OpenRouter yet. Multiple HN commenters checked. No timeline was given.

Getting an API key involves some friction — redirect loops, multiple verification steps. Several people in the thread complained about this.

And the benchmarks compare against Opus-4.6, not Opus-4.7. GPT-5.5 and Gemini 3.1 Pro are missing from most tests. The community noticed. Take the numbers as strong but not definitive.

Who Should Actually Care

If you’re paying $20/month for Claude Code and regularly hitting your weekly limits, Qwen3.7-Max is worth testing. The coding benchmarks are competitive, and the long-horizon agent capabilities are best-in-class.

If you need open-weight models you can run locally, this isn’t it. Watch for the Qwen3.7 open-weight releases — the 122B and 397B variants that HN commenters are already asking about.

If you’re using DeepSeek V4 Pro or GLM 5.1 for agent tasks, Qwen3.7-Max is a meaningful upgrade — especially for long-running autonomous work.

If you’re happy with your current setup and not hitting limits, wait. OpenRouter support will come, and the open-weight models will follow.

Is Alibaba becoming the most interesting player in AI coding tools? I think they might be. Drop a comment with your take.

Qwen, AI Coding Tools, Claude Code, Alibaba Cloud, LLM Benchmarks