N
NexusRoute
Back to Models

Qwen 3.5

Alibabamid

Alibaba's open-weight MoE model with 397B total parameters (17B active). Supports 201 languages — the most multilingual model available. Features a 262K context window and Apache 2.0 licensing. Particularly strong on CJK languages and emerging market languages.

Released 2026-01-28Knowledge cutoff: 2025-10
Medium confidence|Updated 72d ago|85% source confidence

Specifications

Context Window

262.1K tokens

Max Output

32K tokens

Input Price

$0.300 / 1M tokens

Output Price

$1.20 / 1M tokens

Latency Tier

Fast (speed score: 7/10)

Capability Profile

Cost Efficiency
8.5/10
Coding
8/10
Reasoning
7.5/10
Long Context
7.5/10
Structured Output
7.5/10
Instruction Following
7.5/10
Conversational
7.5/10
Speed
7/10
Creativity
7/10
Factuality
7/10
Tool Use
7/10
Multimodal
6.5/10
Safety & Enterprise
6/10

Feature Support

Vision Yes
Audio In No
Audio Out No
Video No
Image Generation No
Image Editing No
Function Calling Yes
JSON Mode Yes
Structured Output Yes
Streaming Yes
Reasoning No
Realtime No
Computer Use No
Web Search No

Best Use Cases

Multilingual applications requiring 200+ language support
Chinese-English bilingual enterprise applications
CJK language processing (Chinese, Japanese, Korean)
Open-weight deployment with Apache 2.0 for commercial use
Cost-effective coding assistance across multiple programming languages
Emerging market applications in Arabic, Hindi, Swahili, etc.

Not Ideal For

Enterprise deployments with Western safety/compliance requirements
Tasks where English-only quality needs to be absolute best
Audio or video processing
Safety-critical applications in regulated Western markets

Strengths

201 language support — the most multilingual model available
Excellent CJK (Chinese, Japanese, Korean) language capabilities
Apache 2.0 licensing for unrestricted commercial use
Strong coding across programming languages — competitive with Llama 4 Maverick
Good mathematical reasoning at its price tier
Vision-language understanding for image+text tasks

Weaknesses

English quality is good but slightly below Western frontier models
Safety alignment reflects Chinese regulatory requirements
Content censorship on Chinese political topics is hard-coded
262K context is respectable but below the 1M+ leaders
Self-hosting the full model requires multi-GPU infrastructure
Vision quality is below Gemini and GPT models

Edge Cases & Notes

17B active parameters per token — efficient inference despite large total size
Best open model for Arabic, Hindi, and Southeast Asian languages
Chinese political censorship applies regardless of the hosting provider
Pricing shown is approximate (Together AI) — varies by provider

Provider Notes

Open-weight under Apache 2.0. Available through Together AI, DashScope (Alibaba Cloud), and self-hosted. The best choice for teams needing broad multilingual support with open weights.

Benchmarks

MMLU87%
HumanEval87.5%
Arena Elo1300

Benchmark Notes

MMLU-Pro 87%. Strong multilingual benchmarks — #1 on many non-English language evaluations. HumanEval 87.5% shows strong coding. Particularly impressive on CMMLU (Chinese) and JMLU (Japanese).

Research Meta

Last Evaluated

2026-03-15

Source Confidence

85%

Evaluation Method

Open LLM Leaderboard, LMSYS Arena, multilingual evaluations (CMMLU, JMLU, ArabicBench), coding benchmarks

Needs Re-evaluation

No

Sources

  • Qwen 3.5 technical report
  • Open LLM Leaderboard
  • LMSYS Chatbot Arena
  • Multilingual benchmark suites

Continue exploring

Route a prompt

See how Qwen 3.5 ranks

Compare models

Side-by-side analysis

Browse registry

Explore all 24 models