FULL STORY

NVIDIA Nemotron 3.5: Released Open-Source, Tops US Open-Weight Charts

NVIDIA released and open-sourced the Nemotron 3.5 family on Oct 15, led by the 120B multimodal MoE Super VL built for long-horizon agents. It quickly topped Artificial Analysis' US open-weight intelligence rankings.

2026-10-15 ~ 2026-10-16 · 3 episodes · 38 posts

Episode 1 · NVIDIA Launches Nemotron 3.5: Super 120B and Lightning 30B Dual Models (2026-10-15, 5 posts)

The NVIDIA Nemotron team has released two new models in the Nemotron 3.5 family: Super and Lightning. Super has 120B total parameters with roughly 12B activated, while Lightning comes in at 30B/3B. Aleksa Gordić, a well-known developer from the GPT/open-source community, took part in the announcement and revealed that he joined the Nemotron pretraining team several months ago and is working on even larger models that are "very big, very strong." The release matches the timeline of earlier leaks, making future model updates worth watching closely.

Confirmed

  • Nemotron 3.5 Super released with 120B total parameters and about 12B activated (m4, m5)
  • Lightning series launched alongside it at 30B/3B scale (m4, m5)
  • Aleksa Gordić joined the NVIDIA Nemotron pretraining team months ago and is helping build larger models (m4, m5)
  • Super features VL (vision-language) capabilities (m5)

Timeline

  • 10-15: testingcatalog leaked that NVIDIA would release Nemotron 3.5 Super VL the following week (possibly at the GTC event on Wednesday the 21st); the model had previously been spotted in official documentation before being pulled, and a link was later posted again as a teaser (m1, m2)
  • 10-15: A forwarded post said Super is compact enough to fit on a single DGX Spark, with multiple models launching the same day alongside benchmark charts such as SWE-bench Multilingual (m3)
  • 10-16: Gordić confirmed the release of Nemotron 3.5 Super and Lightning (m4, m5)

Why it matters

  • The sparse design of 120B total parameters with only 12B activated means a high-performance model can run on a single DGX Spark-class device, dramatically lowering the deployment barrier
  • Gordić's hints at "bigger and stronger" models suggest NVIDIA's continued investment in open-source/open-weight models, poised to shake up the open model landscape

Episode 2 · NVIDIA Open-Sources Nemotron 3.5 Family: 120B Multimodal Super VL and 24GB Bolt (2026-10-15, 31 posts)

On October 15, NVIDIA released and open-sourced the multimodal model Nemotron 3.5 Super VL, which merges vision and text understanding into a single model for long-running Agent workloads, with weights, datasets, and training recipes all made public.

Confirmed

  • Architecture and specs: a MoE with 120B total parameters and 12B active per token, plus a vision encoder; supports text, image, document, and video understanding along with reasoning and tool calling, with up to 1M tokens of context (Artificial Anlys data)
  • Hugging Face releases include base, BF16, and NVFP4 quantized versions plus a LiLiCorr speculative decoding config; per NVIDIA team member gordicaleksa, the open release covers Base BF16 and post-trained BF16/NVFP4 weights, datasets, and training recipes (kuchaev post)
  • Performance: NVIDIA cites Artificial Analysis data showing 640 token/s output speed, roughly 5x faster than open models in the same Intelligence Index band at 100K input tokens; on 4 B200 GPUs, 10,000 SWE-Bench Pro coding tasks completed in about 12 hours with up to 4x speedups; gordicaleksa reports SWE-bench Pro 51.78% and 2.34x throughput
  • Leaderboards: Artificial Analysis gives it an Intelligence Index of 26, topping US open-weights models; kuchaev says it beats their own Ultra-3.0 on most benchmarks
  • Ecosystem: Baseten (Model APIs and Loops) and Together AI offer day 0 support, with Together highlighting NVFP4 KV cache; official demos include a visual investigation Agent that spots store safety hazards, analyzes charts, and audits accounts, runnable on a single DGX Spark
  • NVIDIA also launched the 74B-A9B Nemotron 3.5 Bolt (2-bit quantized to fit local Agents in 24GB of VRAM) and the Lightning model (mentioned by gordicaleksa)
  • NVIDIA showed accuracy gains after post-training across five verticals: document intelligence, cybersecurity, finance, scientific experiments, and travel

Why it matters

  • Unifying text and vision in one model removes the overhead of passing context between separate text and vision models in traditional setups, directly targeting the hot area of multimodal Agent workflows
  • Fully open (weights + data + recipes), high throughput, and single-GPU runnability lower barriers to local and cloud deployment, shaking up the open multimodal model landscape

11 more related posts →

Episode 3 · NVIDIA Nemotron 3.5 Super VL tops US open-weight models (2026-10-15, 2 posts)

Artificial Analysis benchmarks show NVIDIA's Nemotron 3.5 Super VL scoring 26 on the Intelligence Index, becoming the top-ranked US open-weight model, with doubled intelligence and 60% fewer output tokens.