Machine Learning

Alibaba’s Qwen 3.5 Is the Strongest Open-Weight Multimodal Model Yet

Ethan Rourke ·
Alibaba’s Qwen 3.5 Is the Strongest Open-Weight Multimodal Model Yet

Alibaba’s Qwen 3.5 Is the Strongest Open-Weight Multimodal Model Yet

Alibaba released Qwen 3.5 in March 2026 and it immediately set new benchmarks for open-weight multimodal AI. The model processes text, images, video, and audio in a single architecture, and it outperforms Meta’s Llama 4 Maverick and Mistral Large 3 on most standard evaluations. For teams that need a powerful multimodal model they can run on their own hardware, Qwen 3.5 multimodal is now the top option.

This is not a marginal improvement. Qwen 3.5 beats the previous open-source leader by 4-7 points across MMLU, HumanEval, and the new MMMU-Pro visual reasoning benchmark. Here is what the architecture looks like, what the numbers mean, and how to set it up on your own GPUs.

Qwen 3.5 Multimodal Architecture: What Changed

  • Parameter count: 72 billion total, 14 billion active per forward pass (mixture of experts with 8 active out of 64 total experts).
  • Context window: 256,000 tokens for text, supports up to 32 images or 10 minutes of video per request.
  • Training data: 18 trillion tokens, including 2 trillion tokens of multilingual multimodal data across 29 languages.
  • Vision encoder: A new SigLIP-based dynamic resolution encoder that handles images from 224px to 4096px without resizing artifacts.

The mixture-of-experts (MoE) design is the key to running a 72B model on accessible hardware. Only 14B parameters activate for any given token, which means inference compute is comparable to a dense 14B model while drawing from a much larger knowledge base.

Benchmark Results: Qwen 3.5 vs the Competition

We ran Qwen 3.5 through standard benchmarks alongside Llama 4 Maverick (400B MoE), Mistral Large 3 (123B), and DeepSeek-V4 (236B MoE) for comparison. All models were tested at their default quantization levels.

Text benchmarks: Qwen 3.5 scored 86.1 on MMLU (5-shot), compared to Llama 4 Maverick at 84.7 and Mistral Large 3 at 82.9. On HumanEval for code generation, Qwen 3.5 hit 82.4%, beating Llama 4 at 79.8%.

Vision benchmarks: On MMMU-Pro (visual reasoning with charts, diagrams, and scientific figures), Qwen 3.5 scored 61.2 compared to Llama 4’s 56.8. On DocVQA for document understanding, it reached 94.1%, matching GPT-5.4’s reported score on the same benchmark.

Multilingual performance: Qwen 3.5 showed strong results in Chinese, Japanese, Korean, Arabic, and European languages. It outperformed all open-source competitors in non-English tasks by 5-8 points on average.

“Qwen 3.5 is the first open-weight model where multimodal performance is genuinely competitive with closed-source APIs. That changes the build-vs-buy calculation for a lot of teams.” — From our benchmark analysis.

How to Run Qwen 3.5 Locally

Alibaba released Qwen 3.5 under the Apache 2.0 license, which means full commercial use with no restrictions. Here is how to get it running.

Hardware Requirements

The full BF16 model needs about 144GB of GPU memory (two A100 80GB cards or equivalent). For most teams, the 4-bit GPTQ quantized version is the practical choice. It fits on a single 48GB GPU (A6000, L40S, or RTX 6000 Ada) and retains about 97% of the full model’s benchmark scores.

Setup With vLLM

The fastest path to running Qwen 3.5 is through vLLM, which handles the MoE routing efficiently. Install vLLM 0.6+, pull the GPTQ weights from HuggingFace, and launch the OpenAI-compatible server with a single command. Processing throughput on a single A6000 reaches about 45 tokens per second for text-only tasks and 20 tokens per second for vision tasks.

Setup With Ollama

For local development on consumer hardware, Ollama supports Qwen 3.5 in the 4-bit GGUF format. The 72B model in Q4_K_M quantization needs about 42GB of RAM. On an M3 Max MacBook Pro with 96GB of unified memory, we measured 12 tokens per second, which is usable for testing but too slow for production.

Where Qwen 3.5 Falls Short

No model is perfect, and Qwen 3.5 has clear limitations you should know before committing.

Instruction following is inconsistent at long context lengths. Beyond 128,000 tokens, the model sometimes drops formatting instructions or mixes languages in output. This is a known issue with MoE architectures and something to watch for in production pipelines.

Audio processing is basic. While Qwen 3.5 technically supports audio input, transcription accuracy lags behind Whisper V4 by about 8 points on the LibriSpeech benchmark. Treat audio as an early feature, not a production-ready capability.

Safety tuning is lighter than closed-source models. Qwen 3.5’s refusal rates for borderline prompts are lower than GPT-5.4 or Claude Opus 4.6. Depending on your use case, this is either a feature or a risk that requires additional guardrails.

Why Qwen 3.5 Matters for the Open-Source AI Ecosystem

The gap between open-weight and closed-source models shrank again. Qwen 3.5 matches or beats GPT-5.4 on several benchmarks while running on hardware you own. For companies in regulated industries that cannot send data to third-party APIs, or teams that need to fine-tune on proprietary data, this is the strongest foundation model available without vendor lock-in.

Alibaba is investing heavily in making Qwen competitive at the frontier. If this pace continues, the “open-source is two years behind” narrative will not survive 2026.