English

← Back to list

Qwen3.7-Max on Bailian: ¥12/M and China's Top Arena Model

itadol5j

Qwen3.7-Max launches on Bailian

🤖 Qwen3.7-Max goes live on Bailian: ¥12/M input, a Chinese model tops the Arena

Beijing time,May 22, 2026,Qwen3.7-Maxis now officially available on the Alibaba Cloud Bailian platform, callable directly via API:

  • API pricing: ¥12/M input tokens, ¥36/M output tokens
  • Token Plan subscription: subscribers can use the model directly
  • 1M-token context: max output 65,536 tokens
  • No.1 among Chinese models: first among Chinese models on the Arena global blind-test leaderboard

💰 API Pricing

Price comparison

Platform Input price Output price
Alibaba Cloud Bailian ¥12/M tokens (~$2.50) ¥36/M tokens (~$7.50)
OpenRouter $2.50/M tokens $7.50/M tokens

Token Plan subscription

Alibaba Cloud Bailian Token PlanThe subscription service is also liveQwen3.7-Max:

  • Subscribers can use it directly: no extra charge
  • Credits billing: input and output tokens for different models convert to Credits by rule
  • Deduction order: draws from your current seat's monthly quota first → automatically switches over once it's used up

🧠 Core Capabilities

Benchmark results

Benchmark Qwen3.7-Max Opus 4.6 Max DS-V4-Pro Max
Terminal Bench 2.0 69.7 65.4 67.9
SWE-Pro 60.6 57.3 59.0
SWE-Multilingual 78.3 77.5 76.2
SciCode 53.5 51.9 —
GPQA Diamond 92.4 91.3 90.1
HLE 41.4 40.0 37.7
HMMT 2026 Feb 97.1 96.2 95.2
IMOAnswerBench 90.0 86.0 89.8
Apex 44.5 34.5 38.3
MCP-Mark 60.8 56.7 57.1
Kernel Bench L3 1.98x / 96% 2.63x / 98% 1.07x / 54%

Agent capabilities

Benchmark Qwen3.7-Max vs.
MCP-Mark 60.8 GLM 5.1 57.5
MCP-Atlas 76.4 Opus 4.6 75.8
Skillsbench 59.2 Kimi K2.6 56.2
SpreadSheetBench-v1 87.0 First-tier level

🔥 35-Hour Long-Horizon Autonomous Execution

Kernel optimization experiment

Qwen3.7-Maxon ahardware platform it had never seen(the T-Head Zhenwu M890 PPU), completedfully autonomous kernel optimization:

Item Value
Execution time About 35 hours
Kernel evaluations 432 runs
Tool calls 1,158 calls
Final speedup vs. the Triton reference implementation10.0xGeometric mean speedup

vs. other models

Model Speedup
Qwen3.7-Max 10.0x
GLM 5.1 7.3x
Kimi K2.6 5.0x
DeepSeek V4 Pro 3.3x
Qwen3.6-Plus 1.1x

🧑‍💻 Claude Code / OpenClaw integration

Claude Code setup

# Claude Code 配置export ANTHROPIC_MODEL="qwen3.7-max"export ANTHROPIC_BASE_URL=https://dashscope-intl.aliyuncs.com/apps/anthropicexport ANTHROPIC_AUTH_TOKEN=<your_api_key>claude

OpenClaw setup

InOpenClaw's config file, add theBailian providerand setqwen3.7-maxas the default model—that's all it takes.

The preserve_thinking feature

Qwen3.7-Maxsupports thepreserve_thinkingproperty:

  • Preserves the thinking content from all previous turns
  • Designed for agentic tasks
  • Keeps long reasoning chains coherent
  • Recommended to enable in Agent scenarios

📊 YC-Bench Startup Simulation

Simulated startup revenue

OnYC-Bench, simulating one year of startup revenue:

Model Revenue Multiple
Qwen3.7-Max $2.08M 2.0x
Qwen3.6-Plus $1.05M –
Qwen3.5-Plus $352K 5.9x gap

📝 How to Use It

Get an API key

Visit theAlibaba Cloud Bailian console:https://bailian.console.aliyun.com

API calls

Use the model IDqwen3.7-max:

  • OpenAI-compatible protocol:https://dashscope.aliyuncs.com/compatible-mode/v1
  • Anthropic-compatible protocol:https://dashscope-intl.aliyuncs.com/apps/anthropic

Try it in Qwen Studio

Try it online:https://chat.qwen.ai

Summary

Qwen3.7-Max:

  • API pricing: ¥12/M input, ¥36/M output
  • Token Plan subscription: subscribers can use it directly
  • 1M-token context: max output 65,536 tokens
  • No.1 Chinese model on Arena: global LLM blind-test leaderboard
  • AAI score of 57: Artificial Analysis Intelligence Index
  • 35-hour long-horizon execution: 1,158 tool calls, 10x speedup
  • Cross-framework generalization: consistent performance across Claude Code / OpenClaw / Qwen Code

One-line summary: 35 hours, 1,158 tool calls, a 10x speedup—that's Qwen3.7-Max's agentic muscle.

Alibaba Cloud Bailian:https://www.aliyun.com/product/bailian

Qwen website:https://qwen.ai