이 모델의 강점은 무엇인가요?

10兆超の巨大なパラメータ数高度な推論能力の提供 256Kの長いコンテキスト窓

이 모델의 약점은 무엇인가요?

不透明なクローズドライセンスモデルの内部構造が非公開オープンソースではない

어떤 용도에 가장 적합한가요?

複雑な論理思考が必要なタスク大規模文書の解析と要約高度な推論を要する専門業務

アリババ독점

Qwen3.7-Max-Preview

Name: Qwen3.7-Max-Preview
Author: アリババ

이 모델 비교

Qwen3.7-Max-Preview is an inference model developed by Alibaba. Equipped with a large-scale configuration of approximately 10 trillion parameters, it supports a context window of up to 256K.

공유:X Hatena LINE

파라미터

10000.0B

컨텍스트

256K

라이선스

Proprietary

출시일

2026-05-20

API 가격

이 모델의 API 가격 정보는 현재 공개되지 않았습니다

강점

・10兆超の巨大なパラメータ数
・高度な推論能力の提供
・256Kの長いコンテキスト窓

약점

・不透明なクローズドライセンス
・モデルの内部構造が非公開
・オープンソースではない

활용 사례

・複雑な論理思考が必要なタスク
・大規模文書の解析と要約
・高度な推論を要する専門業務

심층 분석

Release Date

May 20, 2026

Alibaba Cloud Summit, Hangzhou

Parameters

Proprietary (undisclosed)

Closed-weight model

Context Window

1,000,000 tokens

First Qwen to break 1M barrier

Max Output

65,536 tokens

Architecture

Decoder-only, extended thinking by default

AA Intelligence Index

56.6

Top 10 globally, highest Chinese model

Chatbot Arena Elo

1,475

Input Price

$2.50/1M tokens

Output Price

$7.50/1M tokens

Cached Input

$0.25/1M tokens

90% cache discount

강점

・Top-ranked Chinese model on Artificial Analysis Intelligence Index (56.6), competitive with Claude Opus 4.7 and Gemini 3.5 Flash
・35-hour autonomous coding demo: 1,158 tool calls, 10x speedup over Triton reference kernel
・Exceptional value: $2.50/$7.50 pricing is half the input cost of GPT-5.5 and Claude Opus 4.7
・Works across multiple agent scaffolds (Claude Code, OpenClaw, Qwen Code) with consistent performance
・Leading reasoning scores: GPQA Diamond 92.4, HMMT Feb 2026 97.1, Apex 44.5

약점

・Proprietary — no open weights available, API-only access
・High abstention rate (48% attempt rate) lowers hallucination but reduces usefulness on edge cases
・Text-only input on API (no vision), despite being an agent-focused model
・No permanent free tier; only limited preview access via chat.qwen.ai
・Vision model capability is limited compared to multimodal competitors

경쟁사 비교

Model	Arena	SWE	GPQA	Price
GPT-5.5	~1500+	~82	93.6	$5.00/$30.00
Claude Opus 4.7	~1490+	~81	~91.3	$5.00/$25.00
Qwen3.7-Max	1475	80.4	92.4	$2.50/$7.50
Gemini 3.5 Flash	~1460+	~78	~90	$1.50/$9.00
DeepSeek V4 Pro	~1440+	~80.6	~88	$1.74/$3.48

개요

Qwen3.7-Max is Alibaba's flagship proprietary reasoning model launched May 20, 2026, designed as an agent foundation for long-horizon autonomous tasks. It scores 56.6 on the AA Intelligence Index (highest Chinese model), features a 1M-token context window, and demonstrated a 35-hour continuous autonomous coding run with over 1,000 tool calls. At $2.50/$7.50 per 1M tokens with a 90% cache discount, it offers frontier-level reasoning at roughly half the cost of GPT-5.5 and Claude Opus 4.7.

벤치마크 및 성능

Qwen3.7-Max posts strong scores across reasoning benchmarks: GPQA Diamond 92.4 (ahead of Claude Opus 4.6 at 91.3), HMMT Feb 2026 97.1 (highest in comparison group), Apex 44.5 (well ahead of DeepSeek V4 Pro at 38.3), SWE-Verified 80.4 (on par with Opus 4.6 Max at 80.8), and HLE 41.4 (narrowly beating Opus 4.6 at 40.0). On Terminal-Bench 2.0-Terminus it scores 69.7 (outperforms DS-V4-Pro Max at 67.9). On agent benchmarks: MCP-Mark 60.8, MCP-Atlas 76.4, Skillsbench 59.2, SpreadSheetBench-v1 87, Kernel Bench L3 1.98x median speedup with 96% win rate.

상세 비교

Sits between Claude Opus 4.7 and Gemini 3.5 Flash on the overall AA Intelligence Index. On reasoning, it matches or exceeds GPT-5.5 on several benchmarks (HMMT 97.1, GPQA 92.4). On coding agents, it is on par with Opus 4.6 Max. Its pricing advantage is significant: $2.50/$7.50 vs $5/$25 for Opus 4.7 and $5/$30 for GPT-5.5. DeepSeek V4 Pro is cheaper but scores lower on reasoning. The main trade-off is the high abstention rate (48% attempt rate), which lowers hallucination but also reduces coverage.

커뮤니티 평가

Strong positive reception at launch. The 35-hour autonomous coding run generated significant buzz. AI Twitter praised the price-to-performance ratio. Some concern about the high abstention rate and text-only API limitation. Integration with Claude Code via Anthropic API protocol was noted as a forward-thinking move. Community comparing it favorably against GPT-5.5 on a cost-adjusted basis.

활용 사례

Best suited for long-horizon agentic coding, multi-step workflow automation, research assistance requiring sustained reasoning over hours, and document-scale context processing. The $0.25 cached input pricing makes it ideal for agent workflows that repeatedly reference the same codebase or document. Not recommended for vision tasks (text-only on API) or for users who need the model to attempt answering all queries (high abstention). For budget-conscious teams needing frontier reasoning, it is the best value among top-tier models.