Qwen Qwen3.5

Overview

Qwen3.5 is Alibaba Cloud's flagship open-weight model, released on February 16, 2026. It is a 397-billion-parameter Mixture-of-Experts model that activates only 17 billion parameters per token — delivering frontier-level performance at a fraction of the compute cost. Unlike previous Qwen generations that required separate Vision-Language variants, Qwen3.5 unifies text, image, and video understanding into a single checkpoint via early fusion.

The release is particularly notable for its agentic capabilities: Qwen3.5 posts strong results across key AI agent benchmarks at launch, including AndroidWorld (66.8), BrowseComp (69.0/78.6), and NOVA-63 (59.1), though performance varies across different evaluation suites. It ships under the Apache 2.0 license, making it freely usable, modifiable, and deployable in commercial products without royalties.

What's New

Unified Multimodal Architecture

Previous Qwen versions required users to choose between a text model and a separate VL (Vision-Language) variant. Qwen3.5 eliminates this split. Vision capabilities are built in through early fusion training on multimodal tokens — not bolted on as a post-hoc adapter — allowing the model to natively reason across text, images, and video in a single pass.

• Image understanding across photos, documents, charts, and UI screenshots (official resolution limits not publicly specified)
• Supports hour-scale video understanding (official docs recommend optimizations for long video; exact maximum duration is not specified)
• Achieves 90.8 on OmniDocBench1.5 (document understanding) — highest of any tested model
• Scores 88.6 on MathVision — highest across GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro

Hybrid Architecture: 8.6×–19× Faster Decoding

The most significant architectural upgrade is the combination of Gated Delta Networks (linear attention) with traditional Gated Attention and sparse MoE routing. Compared to the previous Qwen3-Max generation:

• 8.6× faster decode throughput at 32K context
• 19× faster decode throughput at 256K context

Linear attention scales with sequence length linearly (not quadratically), enabling the 1M token extended context window without the memory explosion typical of standard transformers. This is how Qwen3.5 can process entire codebases, 500-page reports, or multi-hour video transcripts in a single prompt.

Strong Agentic Capabilities

Qwen3.5 is positioned as an "agentic era" model. It natively understands screenshots, identifies UI elements, and executes multi-step workflows across desktop and mobile environments. It combines visual understanding with tool-calling and function-calling capabilities so it can act, not just respond.

Benchmark results at launch:

Expanded Multilingual Support

Qwen3.5 supports 201 languages and dialects — up from 119 in Qwen3. The vocabulary expanded from ~150K to approximately 250K tokens (official config: 248,320), which means:

• 15–25% fewer tokens needed for technical multilingual datasets
• 10–60% cost reduction for multilingual AI code generation and processing
• Better handling of low-resource languages due to more efficient tokenization

Apache 2.0 Open-Weight License

The full 397B model weights are available on HuggingFace under Apache 2.0 — no usage restrictions, no royalties, no commercial approval required. Community quantizations (GGUF format via Unsloth) enable deployment ranging from research setups to production clusters.

Availability & Access

Qwen3.5 is accessible through multiple channels:

• Web UI: chat.qwen.ai — free access with rate limits; no API key required
• API (Qwen3.5-Plus): Available via Alibaba Cloud Model Studio (dashscope.aliyuncs.com); requires an Alibaba Cloud account and API key
• Open weights: Qwen/Qwen3.5-397B-A17B on HuggingFace under Apache 2.0
• OpenAI-compatible API: Drop-in replacement for existing OpenAI SDK workflows — change only the base URL and API key

The API uses the model ID qwen3.5-plus-2026-02-15 on Model Studio.

System Requirements for Local Deployment

Running the full 397B model locally requires substantial hardware. All experts must be loaded into memory even though only 17B activate per token:

Supported inference frameworks: vLLM, SGLang, llama.cpp (GGUF), LM Studio, Ollama, and HuggingFace Transformers (use latest versions; refer to the official quickstart for specific version requirements).

Pricing & Plans

Important: API pricing varies by deployment region (Mainland China vs. international) and context length tier. No free API quota is provided — verify current rates on the Alibaba Cloud Model Studio pricing page before building.

The per-token rate is the same whether thinking mode is enabled or disabled. However, enabling thinking typically generates substantially more output tokens (the <think> block counts toward output), so the total API cost per request increases when thinking is on. Compared to some competing AI chatbots that charge an explicit reasoning surcharge, Qwen3.5's approach keeps the unit price fixed while letting token volume determine the cost.

Best For

• AI agent developers building multi-step automated workflows across desktop, mobile, and web environments
• Teams processing large document collections — SEC filings, entire codebases, multi-hour video transcripts — that require 1M+ token context windows
• Developers building multilingual AI applications serving global audiences across 201 languages
• Organizations wanting frontier-quality AI without vendor lock-in, thanks to Apache 2.0 full model weights
• Researchers and teams requiring strong multimodal document understanding (highest OmniDocBench score) or multimodal math (highest MathVision score)
• Cost-sensitive production workloads where Qwen3.5-Plus API pricing undercuts competing frontier models by 60–70%

FAQ