NVIDIA releases open Nemotron 3.5 Super VL and Bolt models
Confirmed
The takeaway
On Oct 15, 2026 NVIDIA released Nemotron 3.5 Super VL, an open (OpenMDW-1.1) 120B-total / 12B-active hybrid Mamba-Transformer MoE vision-language model with up to 1M tokens of context, for enterprise agents that read documents, charts and web pages.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 15 October 2026
Your AI and this story
- GPT-6 Astra168 days after its cutoff
- Claude Opus 5.5107 days after its cutoff
- Gemini 4 Argon198 days after its cutoff
- Grok 4.7137 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 107 days before it.
Key facts
- Super VL: 120B total / 12B active; Mamba-2–Transformer hybrid MoE with C-RADIO vision encoder and multi-token prediction; context up to 1M; OpenMDW-1.1 licence (Hugging Face card)
- Training: ~26T tokens of pretraining and continued pretraining, ~3.8M post-training samples, RL in ~100 open environments
- VentureBeat: on NVIDIA’s chart, post-training lifted medical-claims accuracy from ~28% to ~72% (Claude Opus 5 ~38%) and security-attack resilience from ~35% to ~76% (GPT-6.1 Sol ~66%); task definitions and sample sizes not published
- Artificial Analysis (via VentureBeat): ~625 output tok/s at 100K input on 4× B200 at an index score of ~26, below GLM-5.3-Flash (42) and Qwen3.8-Flash-Next (40)
- Price on OpenRouter at launch: $0.15 input / $0.50 output per 1M tokens, plus a free tier (nvidia/nemotron-3.5-super-vl)
- Bolt 74B-A9B 24GB: native 2-bit, for RTX Spark laptops (on sale Oct 16), GeForce RTX, DGX Spark; NVIDIA says #1 open model in class on PinchBench, #2 on BrowseComp (no scores given)
- Bolt E4B 4GB for voice assistants and device agents due Oct 23
- Partners customizing: Scale AI, AI21, Mercor, Ramp, Lila Sciences and others; integrating: Atlassian, CrowdStrike, Perplexity, Salesforce, Synopsys
What happened
NVIDIA expanded its open Nemotron 3.5 family on Oct 15, 2026. The main model, Super VL, is a 120B / 12B-active vision-language model for agents that must read tables, charts, PDFs and web pages and call tools. NVIDIA released it with training recipes and data for customization, and showed post-trained variants beating closed frontier models on five narrow tasks. VentureBeat noted that these are customized models on specific tasks, not evidence that the base model beats Claude Opus 5 or GPT-6.1 Sol in general.
The Bolt models target local PCs: a 74B-A9B model compressed natively to 2 bits so it fits in 24 GB, timed for the RTX Spark laptops that go on sale Oct 16, and a 4 GB model due Oct 23.
Why it matters
NVIDIA is now one of the main US sources of open weights, alongside Meta’s Muse Spark Open released the same day. Its pitch is cheap, fast, customizable agents, rather than topping general leaderboards.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 3 primary, 1 press
Primary
- NVIDIA blog (Kari Briski): Open from data center to desktop, Nemotron Super VL and Boltblogs.nvidia.com, official
- NVIDIA Technical Blog: Nemotron 3.5 Super VL powers specialized agentsdeveloper.nvidia.com, official
- Hugging Face: NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4huggingface.co, code
Press
Changes
- Filed. </content> </invoke>
