Qwen3.7-Max Launches: #1 Chinese Model, Top-5 Globally on Intelligence Index

​

Highlights

• Top-5 globally · #1 Chinese model: Artificial Analysis Intelligence Index 56.6, beats Gemini 3.5 Flash (55.3), ranks first among Chinese models
• Agent long-horizon breakthrough: Ran 35 hours autonomously with 1,158 tool calls and 432 kernel evaluations — no human in the loop
• Terminal-Bench 2.0 at 69.7: Continued lead on coding/tool-use benchmarks, Terminal-Bench Hard 50.8% (+6.9)
• 1M context window: Doubled from 256K on the predecessor — native long-task capacity
• 31% token-density gain: Produces 31% more output tokens on the Intelligence Index, denser reasoning per answer
• APIYI official-proxy direct: Listed at $1.7140/$5.1420 per 1M tokens (input/output), identical to Alibaba Cloud’s rates

​

Background

​

Detailed Breakdown

​

Key Features

​

Performance Highlights

​

The 35-Hour Autonomous Run

• Task: Optimize the Extend Attention kernel on a T-Head Zhenwu M890 PPU
• Duration: ~35 hours of continuous autonomous execution
• Process: 1,158 tool calls, 432 kernel evaluations, 5 architectural redesigns iterated
• Result: 10.0× geometric mean speedup over the Triton reference across multiple workloads

​

Technical Specs

​

Practical Use

​

Recommended Scenarios

​

Code Example

from openai import OpenAI

client = OpenAI(
    api_key="your-apiyi-key",
    base_url="https://api.apiyi.com/v1"
)

# Basic chat
response = client.chat.completions.create(
    model="qwen3.7-max",
    messages=[
        {"role": "system", "content": "You are a rigorous senior engineer — back claims with concrete evidence and numbers."},
        {"role": "user", "content": "Explain the difference between Triton and PPU custom kernels for attention compute."}
    ]
)
print(response.choices[0].message.content)

tools = [{
    "type": "function",
    "function": {
        "name": "run_shell",
        "description": "Execute a shell command and return stdout/stderr",
        "parameters": {
            "type": "object",
            "properties": {"cmd": {"type": "string"}},
            "required": ["cmd"]
        }
    }
}]

response = client.chat.completions.create(
    model="qwen3.7-max",
    messages=[
        {"role": "user", "content": "Use ripgrep to find every TODO in the repo, grouped by file."}
    ],
    tools=tools,
    tool_choice="auto"
)

​

Best Practices

• Checkpoint long runs: For 35-hour-class tasks, add stage checkpoints in your business layer (save intermediate artifacts, resumable) — avoid losing work to a single failure.
• Lean into 1M context: Whole-codebase reviews or long-contract diffs work better as a single pass than RAG chunking; budget tokens accordingly.
• Capture the token-density gain: Output tokens are 31% higher than the previous gen — denser reasoning per answer, but budget more output tokens per call.
• Pin snapshots before production: Preview models can roll silently. Use a snapshot ID, not the alias, for production regressions.

​

Pricing & Availability

​

Side-by-side Pricing

​

Stacking the Recharge Bonus

• Top up $100 → ~86% of list
• Top up $300+ → as low as 79% of list (depending on tier; see Recharge Promotions FAQ)

​

Available Groups

​

Summary

1. #1 Chinese composite intelligence: 56.6 on the Intelligence Index reframes “can we use a Chinese model” as “which Chinese model”
2. Hands-off long runs: 35 hours autonomous on a real optimization task makes dev-agent / research-agent product shapes feasible on Chinese models
3. Transparent pricing: APIYI’s list price matches Alibaba Cloud exactly via a fixed 1:7 conversion — stackable recharge bonus pushes effective cost down further