アリババプロプライエタリ

Qwen3.6-Max-Preview

Name: Qwen3.6-Max-Preview
Author: アリババ

このモデルを比較

アリババ開発の高性能MoEモデル。多言語対応と高い推論能力を特徴とする。

シェア:X はてブ LINE

パラメータ

コンテキスト長

262K

ライセンス

プロプライエタリ

リリース日

2026-04-20

日本語性能

✅高品質日本語

多言語対応モデルのうち、日本語処理に優れた性能を持つモデル。

API料金

入力料金（1Mトークンあたり）

$1.3

出力料金（1Mトークンあたり）

課金モード: standard

強み

弱み

活用例

深度分析

AA Intelligence Index

#3 globally, behind GPT-5.4 and Claude Opus 4.7

SWE-bench Pro

~58.4%, top among all evaluated models

Terminal-Bench 2.0

65.4%

Tied with Claude Opus 4.6

GPQA Diamond

88.8%

Strong but trails Gemini 3.1 Pro (94.3%)

Input Price

$1.30/1M

~$1.04 via OpenRouter (20% discount)

Context Window

262K tokens

vs Claude Opus 4.7 and GPT-5.4 at 1M

強み

・Leads six agentic coding benchmarks simultaneously including SWE-bench Pro, SciCode (+10.8 over Plus), and SkillsBench (+9.9 over Plus)
・preserve_thinking feature carries reasoning traces across multi-turn agentic workflows, reducing context loss in complex tool-calling chains
・API compatible with both OpenAI and Anthropic SDK formats — drop-in substitution with a single model string change

弱み

・First closed-weight Qwen flagship breaks three years of open-source tradition; no self-hosting or fine-tuning possible
・Slow output speed (~38–45 tok/s) falls below the median of 62 t/s for comparable reasoning models, creating latency challenges
・262K context window is 4× smaller than Claude Opus 4.7 and GPT-5.4 (both 1M), limiting large-codebase and long-document workflows

競合比較

Model	Arena	SWE	GPQA	Price
Claude Opus 4.6	N/A	80.8%	~90%	$15/$75
GPT-5.4	N/A	88.7%	~92%	$2.50/$10
Qwen3.6-Plus	N/A	78.8%	88.2%	$0.29/$1.65

概要

Qwen3.6-Max-Preview is Alibaba's first proprietary, closed-weight flagship model, released April 20, 2026. Built on a sparse Mixture-of-Experts architecture with an estimated ~1 trillion total parameters, it targets agentic coding, scientific programming, and multi-step tool-calling workflows. The model topped six coding benchmarks at launch — SWE-bench Pro, Terminal-Bench 2.0, SkillsBench, SciCode, QwenClawBench, and QwenWebBench — and ranks #3 globally on the Artificial Analysis Intelligence Index with a score of 52. Its signature innovation is the `preserve_thinking` feature, which retains the model's chain-of-thought reasoning across conversation turns, addressing a critical failure mode in multi-step agent loops where context coherence degrades over sequential tool calls. The closed-weight decision marks a significant strategic pivot for Alibaba, which built its global developer reputation on open-source releases across the entire Qwen family. Two open-weight siblings — Qwen3.6-27B and Qwen3.6-35B-A3B — were released the same week under Apache 2.0, signaling a deliberate tiering strategy: open models for the community, proprietary flagship for monetization. This positions Max-Preview directly against GPT-5.4 and Claude Opus 4.7 at the frontier API tier, while lower-cost options like Qwen3.6-Plus ($0.29/$1.65 per 1M tokens) continue serving cost-sensitive workloads. The model arrives as a 'preview' with explicit acknowledgment from Alibaba that development is ongoing. While its agentic coding scores are genuinely competitive — particularly on SWE-bench Pro and SciCode — it trails Claude Opus 4.6 on SWE-bench Verified (80.8% vs ~73%) and GPT-5.4 on composite benchmarks (BenchLM 89 vs 81). The 262K context window, text-only modality, and below-median output speed are practical constraints. Pricing at $1.30/$7.80 per 1M tokens sits between the budget Plus tier and Western frontier pricing, making it a mid-range option for teams optimizing agentic coding cost-per-quality.

ベンチマーク＆性能

## Benchmark Comparison | Benchmark | Qwen3.6-Max-Preview | Claude Opus 4.6 | GPT-5.4 | Qwen3.6-Plus | |---|---|---|---|---| | AA Intelligence Index | 52 (#3) | #2 | #1 (89 BenchLM) | 39.6 | | SWE-bench Pro | ~58.4% (#1) | 53.4% | 57.7% | 56.6% | | SWE-bench Verified | ~73% | 80.8% | 88.7% | 78.8% | | Terminal-Bench 2.0 | 65.4% | 65.4% (tie) | ~65% | 61.6% | | GPQA Diamond | 88.8% | ~90% | ~92% | 88.2% | | HLE | 28.9% | — | — | 25.7% | | SciCode | 46.9% | — | — | 40.7% | | τ²-Bench Telecom | 95.9% | — | — | 97.7% | | IFBench | 76.6% | — | — | 75.2% | | AA-LCR | 69.7% | — | — | 69.7% | | SuperGPQA | 73.9% | — | — | 71.6% | | AIME 2025 | 93% | — | — | ~95.3% | | QwenWebBench ELO | 1558 (#1) | ~1182 (vs 4.5) | — | 1502 | | Output Speed | ~38–45 tok/s | ~80+ tok/s | ~60+ tok/s | ~80+ tok/s | | TTFT Latency | ~3.3s | ~2.5s | ~2.5s | — | The model's strongest performance areas are agentic coding tasks (SWE-bench Pro #1, SciCode +6.2 over Plus) and frontend generation (QwenWebBench ELO 1558, a 376-point lead over Claude Opus 4.5). It shows solid but not leading performance on core reasoning (GPQA Diamond 88.8%) and trails significantly on SWE-bench Verified (~73% vs 80.8% for Claude and 88.7% for GPT-5.4). The model is notably slow at 38–45 tokens per second, well below the median of 62 t/s for comparable models, which impacts latency-sensitive deployments.

詳細比較

### vs Claude Opus 4.6/4.7 Claude Opus maintains a clear lead on SWE-bench Verified (80.8% vs ~73%) and holds a 1M token context window (4× larger). Terminal-Bench 2.0 is a tie at 65.4%. Claude's MCP Atlas score (77.3% vs ~48% for Qwen Plus) suggests stronger real-world agentic ecosystem integration. However, Qwen3.6-Max-Preview leads on SWE-bench Pro, SciCode, and QwenWebBench by significant margins. At $1.30/$7.80 per 1M tokens vs Claude's $15/$75, Qwen is ~10× cheaper on input and ~10× cheaper on output, making it the clear cost-performance winner for teams that don't need Claude's context length or MCP depth. ### vs GPT-5.4 GPT-5.4 leads on composite benchmarks (BenchLM 89 vs 81) and SWE-bench Verified (88.7% vs ~73%) by a substantial margin. GPT-5.4 offers 1M context and generally faster inference. Qwen3.6-Max-Preview narrows the gap on SWE-bench Pro (58.4% vs 57.7%) and ties on Terminal-Bench 2.0. For teams already embedded in the OpenAI ecosystem, GPT-5.4's broader capability profile and established tooling make switching difficult. For cost-conscious teams running agentic coding pipelines where SWE-bench Pro matters more than SWE-bench Verified, Qwen offers competitive performance at lower cost. ### vs Qwen3.6-Plus Max-Preview delivers meaningful improvements over Plus: +9.9 on SkillsBench, +10.8 on SciCode, +3.8 on Terminal-Bench 2.0, and +2.3 on SuperGPQA. However, Plus offers 4× the context window (1M vs 262K), 2–3× faster output speed, multimodal capabilities, and costs ~7–5× less ($0.29/$1.65 vs $1.30/$7.80). For most production workloads, Plus remains the better value proposition. Max-Preview is justified only when agentic coding performance at the margin is worth the 5–7× cost premium.

コミュニティ評価

The developer and research community reaction has been sharply divided. Practitioners who tested the model on real agentic coding workflows reported that `preserve_thinking` produces noticeably more coherent multi-step tool-calling behavior — the model maintains consistent mental models across sequential tool calls rather than re-deriving understanding from scratch each turn. This was cited as the most practically impactful feature for production agent pipelines. The closed-weight decision generated the most sustained controversy. Alibaba built its AI developer credibility on open releases — nearly one billion downloads across the Qwen series on HuggingFace. The community concern is not about the two open siblings (27B and 35B-A3B remain Apache 2.0) but about whether this signals future Qwen flagships will follow the same closed pattern, undermining the open ecosystem trust built over three years. Benchmark skeptics raised valid concerns. Two of the six claimed '#1' benchmarks (QwenClawBench, QwenWebBench) are Alibaba-internal evaluations. The Terminal-Bench 2.0 claim was presented as a win using Claude Opus 4.5 as baseline, while Opus 4.6 also scores 65.4% — making it a tie, not a victory. One independent reviewer (Towards AI) ran 20 real coding tasks against Max-Preview, Claude Opus 4.7, and GPT-5.4 and concluded the '#1 on 6 benchmarks' headline did not translate to superiority on practical task completion. Hacker News threads flagged instances of the model repeating failed actions without adapting — precisely the behavior `preserve_thinking` is designed to prevent. Adoption patterns show the model gaining traction primarily in agentic coding toolchains: OpenClaw, LangChain, Hermes Agent, and Kilo Code are among the top consumers on OpenRouter. Enterprise adoption remains cautious given the preview status and lack of production SLAs.

ユースケース

### 1. Agentic Coding Pipelines Requiring Multi-Step Tool Chaining When your agent needs to make 10–20 sequential tool calls — navigating codebases, reading files, writing patches, running tests — `preserve_thinking` ensures the model maintains context coherence across the entire chain. This is the model's most differentiated use case. Choose it over Claude Opus or GPT-5.4 when your specific bottleneck is context loss in multi-turn agentic loops, and your tasks fit within 262K context. ### 2. Scientific and Research Code Generation The +10.8 improvement over Qwen3.6-Plus on SciCode makes this the strongest API option for scientific computing tasks: numerical simulation code, data pipeline construction, research-grade Python, and domain-specific scientific programming. Choose Max-Preview over GPT-5.4 or Claude when your workload is specifically scientific code generation rather than general software engineering. ### 3. Frontend Code Generation at Scale QwenWebBench ELO 1558 (vs ~1182 for Claude Opus 4.5) represents a genuine 376-point lead on frontend generation across seven categories: Web Design, Web Apps, Games, SVG, Data Visualization, Animation, and 3D. For teams building vibe-coding pipelines, design-to-code systems, or AI-assisted UI generation, this is the strongest available model on this specific task class. Caveat: this benchmark is Alibaba-internal; validate against your own frontend generation workload before committing. ### 4. Budget-Conscious Teams Needing Frontier Agentic Coding At $1.30/$7.80 per 1M tokens, Max-Preview is ~10× cheaper than Claude Opus and ~2× cheaper than GPT-5.4 on comparable agentic coding tasks. For teams running high-volume coding agents where SWE-bench Pro performance matters and context windows under 262K are sufficient, the cost-per-quality ratio is compelling. Use it when Claude and GPT-5.4 pricing makes your unit economics unviable.