DEVOURED
Nvidia released Nemotron 3.5 Super VL
Nvidia's new 120B-parameter Nemotron 3.5 Super VL model delivers high-throughput multimodal reasoning for specialized agentic workflows.
What: Nvidia released Nemotron 3.5 Super VL, a 120B-parameter multimodal model optimized for tool use and reasoning. Key optimizations include LiLiCorr speculative decoding and NVFP4 quantization, enabling 5x higher throughput for agentic tasks on a single GPU compared to peers.
Why it matters: This indicates a trend where AI labs are optimizing models specifically for 'agentic' workflows—multistep processes involving tool use and multimodal context—rather than just raw text generation.
Takeaway: Developers can access the model via NVIDIA NIM or Hugging Face to implement domain-specific agent post-training using NeMo libraries.
Deep dive
- 120B-parameter mixture-of-experts (MoE) architecture activating 12B parameters per token.
- Integrates reasoning, tool use, and understanding across text, images, documents, and video.
- Optimized for agentic tasks with LiLiCorr speculative decoding (2.3x faster token generation).
- Supports NVFP4 4-bit weights and activations for reduced memory footprint.
- Post-training enabled by NeMo libraries for supervised fine-tuning and reinforcement learning.
- Open datasets provided for tool-use and finance reasoning.
Decoder
- LiLiCorr: A speculative decoding method that jointly evaluates token candidates across a draft to improve generation speed.
- NVFP4: A 4-bit precision format designed for the NVIDIA Blackwell architecture to optimize model size and speed.
- Speculative Decoding: A technique where a smaller, faster model drafts multiple tokens that a larger, primary model verifies in a single forward pass.
Original article
Full article content is not available for inline reading.
Explore all 51 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi