> ## Documentation Index > Fetch the complete documentation index at: https://docs.apiyi.com/llms.txt > Use this file to discover all available pages before exploring further. # Qwen3.7-Max Launches: #1 Chinese Model, Top-5 Globally on Intelligence Index > Alibaba's Qwen team releases flagship Qwen3.7-Max: Artificial Analysis Intelligence Index 56.6 (5th globally, top Chinese model), Terminal-Bench 2.0 69.7, 1M context. APIYI's official-proxy channel is listed at $1.7140/$5.1420 per 1M tokens — identical to Alibaba's rates. ## 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 On **May 20, 2026 (UTC+8)**, Alibaba's Qwen team released its flagship **Qwen3.7-Max**, positioned as "the next-generation flagship model for the agent era." Compared to Qwen3.6 Max Preview, the Artificial Analysis Intelligence Index jumped from 51.8 to **56.6** (+4.8) — pushing a Chinese model into the **global top 5** of this composite leaderboard for the first time, ahead of Google's Gemini 3.5 Flash (55.3). What's more striking is the **agent long-horizon capability**: in an internal test disclosed by Alibaba, Qwen3.7-Max autonomously optimized an Extend Attention kernel on a T-Head Zhenwu M890 PPU. It ran continuously for **35 hours**, executing 1,158 tool calls and 432 kernel evaluations, iterating through 5 distinct architectural redesigns, and finally achieving a **10× geometric mean speedup** over the Triton reference implementation. The entire process was **fully hands-off** — a "let-it-run-a-day-and-a-half" stability that previous Qwen 3.x models did not reach. **Sources**: Qwen official blog `qwen.ai/blog`, Artificial Analysis Intelligence Index 2026/5 data, TechNode coverage (2026/5/21), Digg / Pandaily / SCMP composite reports. Data retrieved 2026/5/21 (UTC+8). ## Detailed Breakdown ### Key Features Artificial Analysis Intelligence Index 56.6, ahead of Gemini 3.5 Flash — first Chinese model in the global top 5. 35-hour continuous run on a single task, 1,158 tool calls, 432 evaluation iterations — no human handholding. Doubled from the previous 256K — natively handles long-document analysis, large codebase reading, multi-hour conversation. Terminal-Bench 2.0 69.7, Terminal-Bench Hard 50.8% (+6.9 vs Qwen3.6 Max Preview) — real-world tool-call workflows lead. ### Performance Highlights Compared to Qwen3.6 Max Preview, the gains concentrate in scientific reasoning, agent capability, and coding: | Benchmark | Qwen3.6 Max Preview | Qwen3.7-Max | Δ | | -------------------------------------- | ------------------- | ----------- | ------- | | Artificial Analysis Intelligence Index | 51.8 | **56.6** | +4.8 | | Terminal-Bench 2.0 | — | **69.7** | — | | Terminal-Bench Hard | 43.9% | **50.8%** | +6.9 pp | | Humanity's Last Exam | 28.9% | **38.1%** | +9.2 pp | | CritPt | 3.7% | **13.4%** | +9.7 pp | | GDPval-AA (Elo) | 1504 | **1546** | +42 | On factuality, Qwen3.7-Max shows "higher abstention" on **AA-Omniscience** — attempt rate falls to 48.0%, the lowest among comparable frontier models. The model is more willing to say "I don't know" rather than confidently produce a wrong answer. For production agent workflows, this self-awareness matters: a wrong answer is more dangerous than silence. ### The 35-Hour Autonomous Run Alibaba's blog details a fully autonomous kernel-optimization case study: * **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 This scale of agent task is difficult on previous models — context blows up, goals drift, errors compound. Qwen3.7-Max gets there via 1M context, drift-resistant instruction following, and stable execution of the "tool-fail → retry → adjust" loop. ### Technical Specs | Spec | Value | | -------------- | ------------------------------------------------- | | Model ID | `qwen3.7-max` | | Context window | 1,000,000 tokens | | Modalities | Text in / text out | | Release status | Preview, API access | | Channel | APIYI official-proxy (Alibaba Cloud Model Studio) | **Preview stage**: Qwen3.7-Max is currently in Preview; open weights have not been released. For production use, follow the usual snapshot-pinning and regression-testing practices for Preview models. ## Practical Use ### Recommended Scenarios Multi-step, cross-tool, long-running tasks (code refactor, research, data pipeline maintenance). Terminal-Bench 2.0 69.7 + Hard 50.8% — suited to code-gen, debug, and tool-heavy workflows. 1M context fits large codebases, long reports, long contracts; reduces RAG chunking overhead. Humanity's Last Exam +9.2, CritPt +9.7 — significant gains on scientific reasoning and open-ended problems. ### Code Example ```python theme={null} 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) ``` Agent workflow (Function Calling) example: ```python theme={null} 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 | Item | Alibaba Cloud (CNY) | APIYI list (USD) | | ------ | ------------------- | -------------------- | | Input | ¥12 / 1M tokens | \$1.7140 / 1M tokens | | Output | ¥36 / 1M tokens | \$5.1420 / 1M tokens | **Exchange rate**: APIYI uses a **fixed 1:7** rate to convert Alibaba's CNY pricing to USD list (12 ÷ 7 ≈ 1.7143; 36 ÷ 7 ≈ 5.1428). This is a fixed conversion, not a preferential rate — it keeps USD-billed customers' line items aligned with the Alibaba Cloud RMB invoice 1-to-1. ### Stacking the Recharge Bonus APIYI runs an always-on [recharge bonus promo](/en/faq/recharge-promotions) — the higher the top-up, the higher the bonus percentage, credited directly to spendable balance: * **Top up \$100 → \~86% of list** * **Top up \$300+ → as low as 79% of list** (depending on tier; see Recharge Promotions FAQ) The discount lives entirely in the bonus credit, separated from the list price. For enterprise volume, reach out via WeChat customer support. ### Available Groups | Group | Open | Notes | | --------- | ---- | ---------------------------------------- | | `Default` | ✅ | Direct calls work on the default group | | `SVIP` | ✅ | High-priority queue, no extra multiplier | ## Summary Qwen3.7-Max isn't just a numbers bump — it pushes **the ceiling on "agent long-horizon capability" for Chinese models** to a new level: 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 **Selection advice**: If your workflow involves **long context + multi-tool calls + multi-step reasoning**, Qwen3.7-Max is the top Chinese-model choice today. For pure chat or single-step tasks, the lighter Qwen3.6-Flash or Qwen3.6-Plus offers better cost-effectiveness. **Sources**: Qwen official blog `qwen.ai/blog`, Artificial Analysis Intelligence Index 2026/5 data, TechNode coverage (2026/5/21), Digg / Pandaily / SCMP composite reports. Data retrieved 2026/5/21 (UTC+8).