AnswersPublished19 sources
What did Nvidia announce on Oct. 15 about its three fully open Nemotron AI agent models—Nemotron 3.5 Super VL, Bolt 74B A9B 24GB and Bolt E4B 4GB—including their parameter counts, memory requirements, quantization, intended uses, availability, coverage from enterprise servers to
15, 2026 announcement expanded Nemotron with Super VL for specialized, always on agents and Bolt models for local AI agent workloads. The available evidence verifies the main launch and benchmark claims, but is insufficient to substantiate all the requested technical details, Navan metrics, integrations and analyst...
Published byImages generated with GPT Image 2
Research answer
Key takeaways
- 15, 2026 announcement expanded Nemotron with Super VL for specialized, always on agents and Bolt models for local AI agent workloads.
- The available evidence verifies the main launch and benchmark claims, but is insufficient to substantiate all the requested technical details, Navan metrics, integrations and analyst observations.
- [7] Models and availability Nemotron 3.5 Super VL: An open, customizable multimodal mixture of experts model with 120 billion total parameters and 12 billion activated per token, intended for agentic workflows involving text and visual info
- [3] Nemotron 3.5 Bolt 74B A9B 24GB: Nvidia described it as a hybrid Mamba–Transformer mixture of experts model for everyday local agent tasks on RTX and DGX Spark, designed to fit into 24GB of memory.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Nvidia announce on Oct. 15 about its three fully open Nemotron AI agent models—Nemotron 3.5 Super VL, Bolt 74B A9B 24GB and Bolt E4. Article summary: 15, 2026 announcement expanded Nemotron with Super VL for specialized, always on agents and Bolt models for local AI agent workloads.. Topic tags: general web, agents, ai, workflow, productivity. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, clickbait thumbnails, icons, and tiny thumbnail layouts. Make it
Nvidia’s Oct. 15, 2026 announcement expanded Nemotron with Super VL for specialized, always-on agents and Bolt models for local AI-agent workloads. The available evidence verifies the main launch and benchmark claims, but is insufficient to substantiate all the requested technical details, Navan metrics, integrations and analyst observations. 7
Models and availability
- Nemotron 3.5 Super VL: An open, customizable multimodal mixture-of-experts model with 120 billion total parameters and 12 billion activated per token, intended for agentic workflows involving text and visual information.
3
- Nemotron 3.5 Bolt 74B-A9B 24GB: Nvidia described it as a hybrid Mamba–Transformer mixture-of-experts model for everyday local agent tasks on RTX and DGX Spark, designed to fit into 24GB of memory.
13 An earlier tool result also identified native 2-bit quantization.
12
- Availability: Earlier search results reported that Super VL and Bolt 74B-A9B were released on Oct. 15, with Bolt E4B scheduled for Oct. 23.
- Remaining specifications: Insufficient evidence to confirm Bolt E4B’s actual parameter count, quantization and memory footprint, Super VL’s memory requirement, or the complete coverage of laptops and embedded devices. The model names alone should not be treated as verified specifications.
Openness, training and performance
- Training recipes: Nvidia’s documentation describes Super’s training stages as pretraining, supervised fine-tuning, reinforcement learning with verifiable rewards, specialist teachers, multi-teacher on-policy distillation and evaluation.
6 The retrieved evidence does not establish exactly which weights and datasets were released, so it cannot fully substantiate “fully open” in the broader sense requested.
- Training-token total: Nvidia explicitly documented 25 trillion pretraining tokens and NVFP4 training for the earlier Nemotron 3 Super.
- Nvidia’s speed claim: Its Super VL technical-blog summary claims up to five times higher single-GPU throughput for agentic workloads; the launch headline likewise advertises “5x faster.”
7 These are Nvidia’s claims, not a demonstrated universal speed advantage.
- Artificial Analysis: The earlier retrieved assessment ranked Super VL as the leading U.S. open-weight model, scoring 26 on its Intelligence Index versus 13 for Nemotron 3 Super, while using approximately 60% fewer output tokens per task.
16 That ranking is specifically among U.S. open-weight models—not all models worldwide.
- Accuracy and integrations: Insufficient evidence for the precise accuracy comparisons or the specific integrations by Atlassian, CrowdStrike and Salesforce.
Navan’s deployment and reported benefits
- What is verified: Navan reported gaining early access as part of broadening its model portfolio and emphasized self-hosting for greater control over reasoning quality, speed and efficiency.
8
- Evaluation outcome: Navan’s announcement says its fine-tuned Super VL delivered the strongest overall performance in its internal evaluation, with the lowest latency and significant efficiency.
1
- What remains unverified: Insufficient evidence for the three-week training-to-production timeline, the particular NeMo AutoModel recipe used, the names and results of all seven evaluation measures, or numerical reductions in conversation costs and latency. The snippets also do not establish a quantified development-speed or quality improvement.
GPU ecosystem and competitive strategy
- Supported interpretation: Nvidia explicitly positions Bolt as bringing agentic intelligence to local Nvidia systems, with the larger Bolt targeting RTX and DGX Spark.
- Unverified competitive details: Insufficient evidence for Alexander Harrowell’s specific observations about Google and Amazon’s custom chips, developers considering replacing Alibaba’s Qwen 3.6 35B with Bolt, or the precise comparison with low-cost Chinese open-weight models. Those details should not be presented as established facts from the retrieved snippets.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What did Nvidia announce on Oct. 15 about its three fully open Nemotron AI agent models—Nemotron 3.5 Super VL, Bolt 74B A9B 24GB and Bolt E4B 4GB—including their parameter counts, memory requirements, quantization, intended uses, availability, coverage from enterprise servers to"?
15, 2026 announcement expanded Nemotron with Super VL for specialized, always on agents and Bolt models for local AI agent workloads.
What are the key points to validate first?
15, 2026 announcement expanded Nemotron with Super VL for specialized, always on agents and Bolt models for local AI agent workloads. The available evidence verifies the main launch and benchmark claims, but is insufficient to substantiate all the requested technical details, Navan metrics, integrations and analyst observations.
What should I do next in practice?
[7] Models and availability Nemotron 3.5 Super VL: An open, customizable multimodal mixture of experts model with 120 billion total parameters and 12 billion activated per token, intended for agentic workflows involving text and visual info
Sources