이 모델의 강점은 무엇인가요?

Massive 10-trillion scale parameter count Extensive 262K context window Advanced conversational AI capabilities

이 모델의 약점은 무엇인가요?

Proprietary model weights are not public Restrictive closed-source licensing Lack of detailed benchmark results

어떤 용도에 가장 적합한가요?

Analysis of ultra-long documents Building complex conversational AI services Advanced processing of large-scale data

모델 목록으로

Alibaba독점

Qwen3.6-Max-Preview

Name: Qwen3.6-Max-Preview
Price: 1.3 USD
Author: Alibaba

Qwen3.6-Max-Preview is a large-scale foundation model developed by Alibaba. With an extremely high computational capacity of approximately 10 trillion parameters, it is a chat-focused model supporting a maximum context window of 262K tokens.

파라미터

10000.0B

컨텍스트

262K

라이선스

Proprietary

출시일

2026-04-20

API 가격

입력 가격 (1M 토큰당)

$1.3

출력 가격 (1M 토큰당)

과금 모드: standard

강점

・Massive 10-trillion scale parameter count
・Extensive 262K context window
・Advanced conversational AI capabilities

약점

・Proprietary model weights are not public
・Restrictive closed-source licensing
・Lack of detailed benchmark results

활용 사례

・Analysis of ultra-long documents
・Building complex conversational AI services
・Advanced processing of large-scale data

심층 분석

Artificial Analysis Intelligence Index

#2 overall out of 201 models as of April 2026

SWE-Bench Pro

57.3%

Top score on this benchmark, per Alibaba claims

Terminal-Bench 2.0

65.4%

Ties with Claude Opus 4.6

Input Price

$6/1M tokens

Standard pricing via Alibaba Cloud; free during preview period

Context Window

262K tokens

Large, but below 1M of some competitors like Qwen3.6-Plus

Total Parameters

>1 trillion

MoE architecture; active parameters not disclosed

강점

・Exceptional agentic coding performance, topping six major benchmarks including SWE-bench Pro and Terminal-Bench 2.0.
・Competitive pricing relative to Western frontier models, with significant cost savings.
・Advanced features like preserve_thinking for reliable multi-step agent workflows and improved instruction following.

약점

・Closed weights and API-only access, preventing self-hosting and limiting transparency.
・Preview model with no production SLA, subject to changes in behavior and pricing.
・Text-only modality at launch, lacking image input support compared to some competitors.

경쟁사 비교

Model	Arena	SWE	GPQA	Price
GPT-5.4	89 (BenchLM composite)	57.7% (SWE-Bench Pro)	N/A	$2.50/$14.00
Claude Opus 4.6	N/A	80.8% (SWE-Bench Verified)	~90%	$15.00/$75.00
Kimi K2.6	N/A	80.2% (SWE-Bench Verified)	N/A	$0.60/$3.00

개요

Qwen3.6-Max-Preview is Alibaba's flagship proprietary large language model, released as an early preview on April 20, 2026. With an estimated parameter count exceeding 1 trillion using a sparse Mixture-of-Experts (MoE) architecture, it is optimized for agentic coding, world knowledge, and instruction following. The model achieves top scores on multiple coding benchmarks, including SWE-bench Pro and Terminal-Bench 2.0, positioning it as a strong competitor to Western frontier models like GPT-5.4 and Claude Opus 4.6. A key innovation is its focus on agentic reliability, with features like preserve_thinking that maintain reasoning context across multi-turn conversations for improved tool-calling and code generation. However, its closed-weight, API-only nature marks a strategic shift for Alibaba, which has historically favored open-source releases. As a preview, it lacks production SLAs and stable pricing, but offers a cost-effective alternative for developers, with input pricing significantly lower than competitors during the preview period.

벤치마크 및 성능

Qwen3.6-Max-Preview delivers significant improvements over its predecessor, Qwen3.6-Plus, across various benchmarks. Official releases highlight gains in agentic coding, world knowledge, and instruction following. Key benchmark scores are compared below: | Benchmark | Qwen3.6-Max-Preview | Qwen3.6-Plus | Improvement | |-----------|---------------------|--------------|-------------| | SkillsBench | #1 (exact score TBD) | Baseline | +9.9 points | | SciCode | #1 (exact score TBD) | Baseline | +6.3 points | | Terminal-Bench 2.0 | 65.4% | 61.6% | +3.8 points | | NL2Repo | +5.0 over Plus | Baseline | +5.0 points | | SuperGPQA | +2.3 over Plus | Baseline | +2.3 points | | QwenChineseBench | +5.3 over Plus | Baseline | +5.3 points | | ToolcallFormatIFBench | +2.8 over Plus | Baseline | +2.8 points | Additional third-party benchmarks from sources like TechSifted and PrimeAIcenter: - SWE-Bench Pro: 57.3% (claimed #1) - Terminal-Bench 2.0: 65.4% (ties Claude Opus 4.6) - QwenWebBench ELO: 1558 (vs. Claude Opus 4.5: ~1182) - BenchLM Composite Score: 81 (vs. GPT-5.4: 89) - Artificial Analysis Intelligence Index: 52 (ranked #2 overall) The model excels in coding-specific tasks but shows gaps in broader composite benchmarks compared to GPT-5.4.

상세 비교

### Qwen3.6-Max-Preview vs. GPT-5.4 - **Pricing**: Qwen3.6-Max-Preview is cheaper at $6/1M input tokens vs. GPT-5.4's $2.50/1M, but output pricing is higher at $24/1M vs. $14/1M. During preview, Qwen may be free. - **Context Window**: Qwen3.6-Max-Preview has 262K tokens; GPT-5.4 offers up to 400K tokens. - **Strengths**: Qwen3.6-Max-Preview leads on SWE-bench Pro (57.3% vs. 57.7% – very close) and offers better cost efficiency. GPT-5.4 has a higher BenchLM composite score (89 vs. 81) and likely better broad reasoning. - **Weaknesses**: Qwen is closed-weight and preview-only; GPT-5.4 is a mature product with proven reliability. ### Qwen3.6-Max-Preview vs. Claude Opus 4.6 - **Pricing**: Qwen is significantly cheaper at $6/$24 vs. Claude's $15/$75 per million tokens. - **Context Window**: Qwen has 262K tokens; Claude supports up to 1M tokens. - **Strengths**: Qwen ties on Terminal-Bench 2.0 (65.4%) and may outperform on specific coding tasks like QwenWebBench. Claude leads on SWE-Bench Verified (80.8% vs. pending for Qwen) and has better multimodal support. - **Weaknesses**: Qwen is text-only and preview; Claude has established enterprise features. ### Qwen3.6-Max-Preview vs. Kimi K2.6 - **Pricing**: Qwen is more expensive at $6/$24 vs. Kimi's $0.60/$3.00, but Kimi is open-weight and can be self-hosted for zero API cost. - **Context Window**: Both have 262K tokens. - **Strengths**: Qwen has stronger official benchmark claims on some coding tasks; Kimi supports native multimodal capabilities and agent swarm architecture for parallel tasks. - **Weaknesses**: Qwen is closed-weight; Kimi's open-weight nature requires significant hardware for self-hosting.

커뮤니티 평가

Developers and researchers have reacted positively to Qwen3.6-Max-Preview's performance, noting its competitive edge in coding benchmarks. From the web search results: - On Reddit and Twitter, users highlight its affordability: "At $1.25 per million tokens, Qwen is significantly cheaper than Claude for large scale data ingestion." (Source: Automatio) - Benchmarks are praised: "Qwen3.6-Max-Preview just beat Claude Opus 4.5 on SWE-Bench Pro. China is catching up fast." (Source: Twitter, via Automatio) - The thinking mode feature is appreciated: "The fact that Thinking Mode is baked in as the default state is a meaningful design choice for agentic reliability." (Source: Twitter, via Automatio) - Some caution exists due to its preview status and closed nature, but it is seen as a strong option for agentic coding workflows. Overall, the community views it as a major step forward for Chinese AI models, with adoption likely in cost-sensitive and coding-centric use cases.

활용 사례

1. **Autonomous Software Engineering**: Ideal for AI agents that need to navigate and modify large codebases. Example: Using Qwen3.6-Max-Preview to automatically fix bugs across multiple files in a GitHub repository, leveraging its top SWE-bench Pro score. Choose this over alternatives when cost is a factor and coding performance is critical. 2. **Scientific and Research Code Generation**: Suited for generating code for data pipelines, simulations, or computational tasks. Example: Building Python scripts for scientific computing, where its +10.8 point improvement on SciCode over Qwen3.6-Plus is beneficial. Prefer this when working on research-grade programming without multimodal needs. 3. **Front-End Code Generation at Scale**: Strong for creating web UIs, dashboards, or interactive applications. Example: Generating React components or SVG visualizations, where its QwenWebBench ELO of 1558 outperforms Claude. Select this for design-to-code pipelines where visual output quality is paramount. 4. **Tool-Calling Agents for Automation**: Reliable for workflows requiring structured function calls and multi-step reasoning. Example: Deploying MCP-based agents for automated data analysis, benefiting from its improved ToolcallFormatIFBench score. Use this over models like GPT-5.4 when prioritizing cost-effective, high-reliability tool execution in agentic loops.