AI News Feed
Market watch
Intero Biosystems Wins Startup Battlefield at TechCrunch Disrupt 2026; Apate Named Runner-UpShield AI Expects First Full-Scale X-BAT VTOL Combat Drone by End of OctoberSurvey: SK hynix Tops List of Most Sought-After Employers Among South Korean Job SeekersCal AI Founder Credits Paid Influencers for Viral GrowthTechRadar Publishes Quordle, NYT Connections and Strands Answers for October 16Withings to Charge New Users $100 a Year for Apple Health Sync
Intero Biosystems Wins Startup Battlefield at TechCrunch Disrupt 2026; Apate Named Runner-UpShield AI Expects First Full-Scale X-BAT VTOL Combat Drone by End of OctoberSurvey: SK hynix Tops List of Most Sought-After Employers Among South Korean Job SeekersCal AI Founder Credits Paid Influencers for Viral GrowthTechRadar Publishes Quordle, NYT Connections and Strands Answers for October 16Withings to Charge New Users $100 a Year for Apple Health Sync
Large Language Models

NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model

NVIDIA released Nemotron 3.5 Super VL, a 120B-parameter open multimodal agent model that activates 12B parameters per token for enterprise agents, plus two edge models.

AI News Feed · 2026-10-16 08:39

According to MarkTechPost, one open checkpoint replaces a separate text model and a separate vision model. Nemotron 3.5 Super VL reads text, images, documents and video, then reasons, calls tools and writes code in one model. It supports up to 1 million tokens of context. The model runs on one DGX Spark, one H200, two H100s or one GB200 with the NVFP4 checkpoint; a BF16 checkpoint is also published.

The weights ship under the OpenMDW-1.1 license, and NVIDIA also released open datasets, training recipes and reinforcement learning environments in NeMo Gym, MarkTechPost reported. The model card lists text, image and video inputs with text output. Video is limited to MP4 clips of up to two minutes, and reasoning can be toggled on or off through the chat template.

The backbone interleaves Mamba-2, mixture-of-experts and attention layers. Each MoE layer holds 512 routed experts, and a C-RADIO vision encoder and connector feed images, document pages and video frames into the language backbone.

NVIDIA attributes speed to four techniques. Multi-token prediction uses a native drafter to propose several tokens that are verified in one forward pass. LiLiCorr is a speculative drafter for DGX Spark that NVIDIA says generates tokens up to 2.3 times faster. NVFP4 with quantization-aware distillation uses 4-bit weights and activations on Blackwell and 4-bit weights with 16-bit activations elsewhere. Context reuse uses prefix caching to keep both attention caches and Mamba states across agent turns.

The model was distilled from Nemotron 3 Ultra. NVIDIA states it trained on about 26 trillion tokens, used roughly 3.8 million samples plus reinforcement learning across dozens of environments in post-training, and added 10 billion Indic tokens in continued pretraining.

All benchmark scores in the report are NVIDIA-reported and come from the NVFP4 checkpoint on the model card. The scores include 89.90 on GPQA-Diamond, 71.07 on SWE-bench Verified, 60.80 on Terminal-Bench 2.1, 77.5 on BrowseComp, 71.50 on IFBench, 79.88 on CharXiv reasoning, 62.82 on MMLongBench-Doc, 52.3 on OSWorld-Verified and 17.78 on tau3-bench banking. NVIDIA also claims up to 5 times higher single-GPU throughput than open peers.

On agent harness evaluations, Super VL beats Nemotron 3 Ultra, scoring 63.44 percent on Hermes Bench versus 56.33 percent and 87.05 percent on LangChain evaluations versus 84.3 percent, according to MarkTechPost. Ultra carries about 4.6 times more total parameters. NVIDIA reports over 600 output tokens per second on Artificial Analysis agentic tasks via a Baseten endpoint, and its chart marks a 5 times speed gap over Inkling Small.