NVIDIA Nemotron 3.5 Bolt 74B Launches with Open License

NVIDIA has officially launched Nemotron 3.5 Bolt, a 74 billion parameter multimodal model released under an open license on October 15, 2026. The A9B 24GB configuration marks a significant shift in NVIDIA's AI strategy, bringing enterprise-grade multimodal capabilities to developers and researchers through the openmdw licensing framework. This release positions NVIDIA directly against closed multimodal systems while maintaining the compute efficiency the company is known for.

Model Architecture and Specifications

Nemotron 3.5 Bolt operates with 74 billion parameters in an A9B architecture optimized for 24GB memory footprints. This configuration enables deployment on high-end consumer GPUs and entry-level data center hardware, significantly lowering the barrier to multimodal AI experimentation. The model processes both visual and textual inputs, handling image analysis, document understanding, and cross-modal reasoning tasks within a unified architecture.

According to NVIDIA's technical documentation, the A9B variant employs attention mechanisms designed for memory-constrained environments without sacrificing inference quality. This allows organizations with modest GPU infrastructure to run production workloads that previously required multi-GPU clusters or cloud-based solutions.

Release Date and Availability

Officially launched on October 15, 2026, Nemotron 3.5 Bolt is immediately available for download and deployment. NVIDIA published model weights, inference code, and integration guides through standard AI research channels. The openmdw license permits commercial use, modification, and redistribution, removing common legal barriers that have limited adoption of other foundation models.

Developers can access Nemotron 3.5 Bolt through NVIDIA's AI catalog and third-party model repositories. The company provides optimized inference runtimes for both Linux and Windows environments, with native support for TensorRT acceleration on RTX 40-series and later GPUs.

Open Licensing Strategy

The openmdw license represents a notable departure from NVIDIA's traditional enterprise-focused licensing models. Key provisions include:

• Unrestricted commercial deployment without royalty payments
• Permission to create derivative works and fine-tuned variants
• Redistribution rights for modified model versions
• Attribution requirements that maintain NVIDIA's brand association

This licensing approach aligns Nemotron 3.5 Bolt with community-driven AI development while preserving NVIDIA's hardware ecosystem advantages. Organizations adopting the model gain flexibility typically associated with fully open source projects, yet benefit from NVIDIA's optimization work and ongoing support.

Competitive Positioning

Nemotron 3.5 Bolt enters a crowded multimodal landscape dominated by GPT-4o, Claude 3.5, and Gemini variants. NVIDIA's differentiation rests on three pillars: open licensing, efficient memory use, and tight integration with the company's GPU stack. While competitors offer API-based access with usage fees, Nemotron 3.5 Bolt allows organizations to self-host and customize without ongoing vendor costs.

The 24GB memory requirement positions the model between smaller open alternatives like LLaVA variants and larger proprietary systems requiring distributed inference. Early benchmarks suggest performance competitive with mid-tier commercial offerings, though NVIDIA has not published comprehensive evaluation results against established multimodal benchmarks.

What This Means

NVIDIA's launch of Nemotron 3.5 Bolt signals increasing competition in the open multimodal AI space. By offering a commercially permissive 74B parameter model optimized for accessible hardware, NVIDIA lowers adoption costs for enterprises hesitant to commit to cloud-based AI services. The openmdw license encourages ecosystem development while keeping deployments tethered to NVIDIA silicon. For developers and researchers, Nemotron 3.5 Bolt provides a new baseline for multimodal experiments without the constraints of proprietary APIs or restrictive licenses. As organizations evaluate AI infrastructure investments heading into 2027, self-hosted open models like Nemotron 3.5 Bolt present a viable alternative to vendor lock-in.