NVIDIA Corporation has released Nemotron-3.5-Bolt-74B, a new vision-language model (VLM) designed to facilitate local execution of agentic tasks. The model utilizes a hybrid Mamba-2–Transformer Latent Mixture-of-Experts (MoE) architecture, incorporating a C-RADIO v4-H vision encoder.
Technical Specifications and Efficiency
The model is characterized by its 74 billion total parameters, though it utilizes only 9 billion active parameters during processing. To facilitate local deployment on consumer-grade hardware, NVIDIA has implemented native 2-bit quantization (IQ2) via quantization-aware training. This allows the model to fit within a 24GB memory footprint, enabling it to run alongside other active applications.
According to technical documentation provided by NVIDIA, the model aims to balance intelligence and speed, positioning itself as faster than most popular dense models while maintaining higher intelligence levels than MoE models with similar memory requirements. The model supports a context length of up to 256,000 tokens and can process text, images, and video as inputs, producing text as output.
Deployment is optimized for NVIDIA Ampere (A100), Hopper (H100-80GB), or Blackwell architectures using vLLM. The release includes support for speculative decoding via the LiLiCorr drafter through DFlash, utilizing seven speculative tokens in companion Bolt vLLM recipes.
Use Cases and Licensing
NVIDIA has positioned the Nemotron-3.5-Bolt-74B for “everyday agentic tasks” performed locally. The model is released under the OpenMDW-1.1 license, making it available for both commercial and non-commercial use.
The model includes a reasoning mode controlled via chat templates, allowing users to enable or disable “thinking” capabilities, a feature inherited from the Super 3.5 VL client. For optimal text generation, NVIDIA recommends a temperature setting of 1.0 and a top_p of 0.95.
Performance and Community Observations
While NVIDIA’s official documentation highlights the model’s efficiency, some unverified reports from community discussions on Reddit suggest a trade-off between intelligence and speed. One user noted that while the model may be listed as less intelligent than certain competitors, such as Qwen 3.8 27B, it is intended to offer significantly higher speeds.
The model’s post-training data has a cutoff date of May 2026. As with many large-scale AI developments, the model’s performance in real-world scenarios remains a subject of ongoing testing by developers and researchers within the open-source community.