Qwen 3.5

Open Source

Alibaba (Qwen)•Released on 2026-02-14

Alibaba's flagship open-source MoE model with 397B total parameters (17B active per pass). Apache 2.0 licensed for commercial use. Supports 201 languages with native vision capabilities. Best open-weight model for local deployment.

84
Overall Score

Core Specs

262K
Context Window
32K
Max Output
✓
Reasoning
✓
Open Source
Multimodal Support
textimage

Scenario Scores

Coding
#8 / 1487
Writing
#9 / 1482
Translation
#3 / 1490
Data Analysis
#9 / 1483
Conversation
#10 / 1480
Image Gen
—

User Feedback Highlights

Based on community feedback. Hover to see original reviews.

− Smaller context than GPT-5.4/Opus 4.6− Requires significant VRAM for local hosting+ Open source (Apache 2.0)+ MoE efficiency (17B active)+ Strong vision/multimodal performance+ Cheapest API among frontier-class+ 201 language support+ Self-hostable with vLLM− Weaker on hard coding tasks vs Opus/GPT− Quantization affects complex reasoning
Sentiment:👍 62%😐 22%👎 16%

Pros & Cons

Pros

  • +Open source (Apache 2.0)
  • +Self-hostable with vLLM
  • +201 language support
  • +MoE efficiency (17B active)
  • +Cheapest API among frontier-class
  • +Strong vision/multimodal performance

Cons

  • −Weaker on hard coding tasks vs Opus/GPT
  • −Requires significant VRAM for local hosting
  • −Quantization affects complex reasoning
  • −Smaller context than GPT-5.4/Opus 4.6

Reliability

SLA99.5%
Incidents (30d)0
支持自托管，可避免 API 依赖。阿里云 API 目前无公开重大故障记录。
View Status Page →

Pricing

Input (per 1M tokens)$0.39
Output (per 1M tokens)$1.56
Free trial available
Updated on 2026-03-06

Tools Supporting This Model

Cline
Open Source

Compare with Others

Qwen 3.5 vs GPT-5.4 →Qwen 3.5 vs GPT-5.4 Thinking →Qwen 3.5 vs Claude Opus 4.6 →

Benchmarks

liveCodeBenchV683.6%
aime2691.3%
videoMme84.5%
mmlu89.2%