Alibaba (Qwen)オープンソース

Qwen3.5-27B

Name: Qwen3.5-27B
Author: Alibaba (Qwen)

Alibaba通義のQwen3.5シリーズ初のDenseモデル。27Bパラメータで1010Kコンテキストをサポート。ハイブリッド推論（思考モード対応）、ネイティブマルチモーダル能力を搭載。Agent能力でGPT-5 miniを、視覚理解でQwen3-VL旗艦モデルとClaude Sonnet 4.5を超越。単一GPUで動作可能。2026年2月25日リリース。

シェア:X はてブ LINE

パラメータ

27B (Dense)

コンテキスト長

1010K

ライセンス

Qwen License

リリース日

2026-02-25

API料金

このモデルのAPI料金情報は現在未公開です

強み

弱み

活用例

深度分析

Arena Code Elo

1358

Strong code-model ranking among open models

GPQA Diamond

85.5%

Graduate-level science reasoning

SWE-bench Verified

72.4%

Software engineering

Context Window

262K tokens

260K input / 65K output

Parameters

27B (dense)

Linear-attention dense model, not MoE

Input Price (OpenRouter)

$0.30/1M tokens

Alibaba direct: $0.086/1M (≤128K)

強み

・Dense 27B linear-attention architecture delivers strong reasoning at a fraction of MoE model size and cost.
・Long 262K context window (260K in / 65K out) suits document-heavy and agentic workflows.
・Open-source under Qwen License, with native image and video understanding feeding into text.
・Competitive with much larger Qwen3.5-122B-A10B on reasoning and coding benchmarks.

弱み

・Trails top closed-source flagships on hardest agentic and math tiers (HLE 48.5%, Terminal-Bench 2.0 41.6%).
・Linear-attention design is newer and less battle-tested than standard transformers for some long-context edge cases.
・Multimodal is understanding-only (image/video → text); no image or video generation.

競合比較

Model	Arena	SWE	GPQA	Price
Qwen3.5-27B	N/A	72.4%	85.5%	$0.30/$2.40
Qwen3.5-122B-A10B	N/A	N/A	N/A	N/A
DeepSeek V3.2	1425	73.10%	82.40%	$0.28/$0.42
Claude Opus 4.7	N/A	N/A	87.30%	$15/$75

概要

Qwen3.5-27B is Alibaba's 27-billion-parameter dense reasoning model released on February 24–25, 2026. Unlike most efficient models that rely on Mixture-of-Experts, it uses a dense linear-attention architecture that keeps the full 27B active while cutting memory and compute versus standard transformers. It pairs a 262K-token context window (260K in / 65K out) with native multimodal understanding (image and video into text), positioning it as a cost-effective open-weight workhorse for reasoning, coding, and long-document agentic tasks. On reasoning and coding benchmarks it rivals the far larger Qwen3.5-122B-A10B, making it one of the most efficient open models in its class.

ベンチマーク＆性能

Qwen3.5-27B posts competitive scores across reasoning, coding, and multimodal understanding: | Benchmark | Score | Notes | |-----------|-------|-------| | GPQA Diamond | 85.5% | Graduate-level science | | HLE | 48.5% | Frontier-level interdisciplinary | | MMLU-Pro | 86.1% | General knowledge | | MMMLU | 85.9% | Multilingual MMLU | | C-Eval | 90.5% | Chinese exams | | MMMU-Pro | 75.0% | Multimodal understanding | | SWE-bench Verified | 72.4% | Software engineering | | LiveCodeBench v6 | 80.7% | Coding | | Arena Code Elo | 1358 | Code-model ranking | | τ²-Bench | 79.0 | Agent tool use | | OSWorld-Verified | 56.2 | Computer use | | Terminal-Bench 2.0 | 41.6 | Terminal agent | | IFBench | 76.5 | Instruction following | It roughly matches the larger Qwen3.5-122B-A10B on reasoning and coding while running at a small fraction of the parameter and serving cost.

詳細比較

**Qwen3.5-27B vs Qwen3.5-122B-A10B:** - **Size:** 27B dense vs 122B-A10B MoE (10B active); the 27B is far cheaper to serve. - **Performance:** Near parity on GPQA, SWE-bench, and LiveCodeBench; the 122B pulls ahead on the hardest tiers. - **Best for:** 27B for cost-sensitive production; 122B for max accuracy. **Qwen3.5-27B vs DeepSeek V3.2:** - **Architecture:** Dense linear-attention (Qwen) vs MoE (DeepSeek). - **Modality:** Qwen adds image/video understanding; DeepSeek is text-only. - **Pricing:** Comparable at ~$0.30/$2.40 vs $0.28/$0.42 per 1M tokens. - **Benchmarks:** Qwen leads on GPQA (85.5 vs 82.4) and SWE-bench (72.4 vs 73.1 nearly even); DeepSeek edges Arena Elo (1425 vs code Elo 1358, different scales). **Qwen3.5-27B vs Claude Opus 4.7:** - **Cost:** Qwen is roughly 50x cheaper per token. - **Ceiling:** Opus leads on hardest agentic/math tiers (HLE, Terminal-Bench). - **Modality:** Both handle multimodal input; Opus adds richer tool/computer use.

コミュニティ評価

Developers welcomed Qwen3.5-27B as a rare efficient dense model that does not sacrifice much reasoning quality. On OpenRouter it quickly became a popular default for self-hosted and cost-sensitive deployments thanks to the $0.30/$2.40 pricing and open Qwen License. Common use cases shared in the community include local coding assistants, long-document RAG pipelines, and multimodal triage (image/video in, structured text out). Some note the linear-attention architecture behaves differently from standard transformers on very long contexts, advising testing before production.

ユースケース

1. **Cost-Sensitive Production Coding:** Use as an in-IDE or API coding assistant where per-token cost matters; LiveCodeBench v6 80.7% and SWE-bench 72.4% cover daily refactoring and bug fixing at a fraction of frontier-model price. 2. **Long-Document RAG and Analysis:** The 262K window (260K in / 65K out) handles entire reports, contracts, or repos in one pass—ideal for summarization, extraction, and Q&A over large files. 3. **Multimodal Triage:** Feed screenshots or short video frames and get structured text (captions, extraction, classification); pairs well with downstream agents that act on the output. 4. **Self-Hosted Edge Deployment:** At 27B dense it fits on a single modern GPU, making it a strong open-weight choice for on-premise or privacy-sensitive workloads that cannot use closed APIs.