Jessie A Ellis Feb 27, 2026 18:05

NVIDIA provides free GPU-accelerated APIs for Alibaba’s 397B parameter model Qwen3.5, which allows developers to create multimodal AI agents.

NVIDIA released free GPU-accelerated ends for Alibaba's Qwen3.5 model vision-language, allowing developers to access the 397 trillion parameter system via Blackwell architecture hardware. Both tech giants are now positioned to capture the growing market for multimodal AI agents that can understand and navigate user interfaces.

Alibaba's Qwen3.5 language model was released on 16 February 2026. It represents a major architectural shift for large language models. The Qwen3.5 model, which Alibaba released on February 16, 2026, represents a significant architectural shift in large language models. Only 17 billion parameters are activated per forward pass despite its 397B total parameter count. This 4.28% activation is achieved by a hybrid mixture of experts (MoE), combined with Gated Delta Networks. Alibaba claims that this efficiency results in real cost savings. The system is 60% cheaper than its predecessor and can handle large workloads 8 times more efficiently.

Noteworthy Technical Specifications

It can process two hours of native video content. It supports 200+ languages, and has 512 experts running per layer. Each token is activated by 11 experts (11 routed plus one shared).

Qwen3.5 is available to developers through NVIDIA build.nvidia.com, after registering for free in the NVIDIA developer program. The OpenAI-compatible API makes integration easy for teams who already work with similar tool-calling conventions.

Production Deployment Option

NVIDIA NIM offers the containerized microservices model for enterprises that want to move beyond experimentation. These can be deployed on-premises or in cloud environments. The NeMo Framework provides fine-tuning for domain-specific apps. NVIDIA highlights a visual QA tutorial that demonstrates radiological datasets.

Alibaba has expanded the Qwen3.5 product family since its initial release. Alibaba released three new variants on February 24: Qwen3.5-122B-A10B; Qwen3.5-35-2B-A3B and Qwen3.5-27-9B. These offer smaller footprint options to suit different deployment scenarios.

Alibaba, which had a market capitalization of $372 billion on February 27, compared Qwen3.5 to GPT-5.2 and Claude Opus 4.5 as well as Gemini 3 Pro in terms of benchmark performance. Hugging Face Hub, ModelScope and NVIDIA managed endpoints still offer open-weight models for developers that prefer to self-host.

Image Source: Shutterstock