- Qwen 3.5, released February 2026, uses a 397B-parameter MoE architecture that activates only 17B parameters per forward pass, delivering strong performance at lower inference cost.
- The model scores 91.3 on AIME 2026 and beats GPT-5.2 on IFBench (76.5 vs. 75.4) while matching Gemini 3 Pro on SWE-bench Verified.
- Released under the Apache 2.0 license with full commercial use rights, supporting 201 languages and 256K context length.
- Qwen3.5-Omni adds native multimodal capabilities across text, images, audio, and video in a single model.
What Is Qwen?
Qwen is an open-source large language model family developed by Alibaba Cloud’s Qwen team. The latest flagship release, Qwen 3.5, launched on February 16, 2026, alongside Qwen3.5-Plus and the multimodal Qwen3.5-Omni variant. The project has grown from a competitive Chinese LLM into one of the top-performing open-weight models globally.
The model family spans multiple sizes, from a compact 0.8B-parameter variant suitable for edge devices to the flagship 397B-A17B model that competes directly with proprietary models from OpenAI and Google.
Key Features
Sparse MoE architecture. The flagship Qwen3.5-397B-A17B contains 397 billion total parameters but activates only 17 billion per forward pass using a Mixture-of-Experts design. This combination of Gated Delta Networks with sparse MoE delivers inference throughput 8.6x faster than the previous Qwen3-Max at 32K context.
201-language support. Qwen 3.5 supports 201 languages and dialects, a significant expansion from earlier versions. The Omni variant adds speech recognition in 113 languages and speech generation in 36 languages.
256K context window. The model handles up to 256K tokens of context, enabling processing of long documents, extended codebases, and multi-turn conversations without truncation.
Native multimodal processing. Qwen3.5-Omni processes text, images, audio, and video simultaneously within a single pipeline, supporting 10+ hours of audio and 400+ seconds of 720p video at 1 FPS.
Pricing
Qwen models are released under the Apache 2.0 license, making them free to use, modify, and deploy commercially. Self-hosting costs depend on infrastructure. Running the flagship 397B model requires substantial GPU memory, though the smaller variants (9B, 27B) run on consumer hardware.
Alibaba Cloud also offers Qwen models via its Model Studio API, with pay-per-token pricing that undercuts OpenAI and Google on equivalent model tiers.
How It Compares
vs. GPT-5.2: Qwen 3.5 beats GPT-5.2 on IFBench (76.5 vs. 75.4) and MathVision (88.6 vs. 83.0) but trails on SWE-bench Verified (76.4 vs. 80.0). GPT-5.2 remains stronger for complex software engineering tasks. Qwen is free and open-weight; GPT-5.2 is not.
vs. Gemini 3 Pro: Near-parity on SWE-bench Verified (76.4 vs. 76.2). Gemini has tighter integration with Google’s ecosystem. Qwen offers more deployment flexibility as an open-weight model.
vs. Llama 4: Both are open-weight. Qwen 3.5 currently leads on most benchmarks, particularly in multilingual tasks and math reasoning. Llama 4 has stronger community tooling in the English-language ecosystem.
What to Know Before Signing Up
Qwen 3.5 is best suited for developers and organizations that want top-tier LLM performance without vendor lock-in. The Apache 2.0 license imposes no restrictions on commercial deployment. The flagship model requires significant GPU resources for self-hosting, but the 9B and 27B variants offer strong performance-to-cost ratios on mid-range hardware.
The model’s multilingual strength makes it particularly valuable for applications targeting non-English markets. Documentation is thorough but some resources are primarily in Chinese. The rapid release cadence (multiple major versions per year) means staying current requires regular model updates.
