NVIDIA's Nemotron 3.5 model family launch

Most models in the Nemotron 3.5 family lean on the same efficiency logic: a mixture-of-experts design where only a fraction of total parameters activate per token - confirmed for Super VL, Lightning, and Bolt 74B-A9B, though NVIDIA has not published an active-parameter breakdown for Bolt E4B or Ultra. Super VL carries 120 billion total parameters but activates just 12-13 billion per token ^[1], trained on roughly 26 trillion tokens plus 3.8 million reinforcement-learning samples ^[12]. That same efficiency ratio runs down the stack - Lightning ships as 30B total with 3B active, and Bolt 74B-A9B keeps 74B total parameters but only 9B active, squeezed further by native 2-bit quantization so the whole thing fits in 24GB of memory. The training recipe is where the post-training story gets specific: NVIDIA reports fine-tuning lifts of 9-44% on narrow tasks, for example moving security-attack resilience from 35% to 76% accuracy ^[1]. Super VL's hybrid Mamba-2/attention layers are also why the model does not simply drop into any inference stack - vLLM 0.27.1 or newer is a hard requirement, since older serving engines don't have the Mamba-2 kernels compiled in ^[2], a detail several outlets flagged as the most likely practical snag for teams trying to self-host on day one.

NVIDIA's launch materials lean hard on head-to-head wins against frontier closed models, with coverage headlined by claims that post-trained Super VL beats Claude Opus 5 and GPT-6.1 Sol on specialized benchmarks ^[3]. NVIDIA's own figures back that up on narrow tasks - medical insurance claims accuracy rises from 28% to 72% after fine-tuning, ahead of Claude Opus 5's roughly 38% on the same test ^[1]. But that framing only holds on narrowly fine-tuned jobs. On Artificial Analysis's broader Intelligence Index, Super VL scores 26 against Claude Opus 5.5's 58 ^[4]- less than half, a gap that undercuts any read of Nemotron 3.5 as a general frontier competitor. Artificial Analysis's own writeup instead positions it as the leading US open-weight model ^[5], a narrower and more defensible claim than the benchmark slides suggest. The takeaway: Nemotron 3.5 is built to win the specific jobs enterprises fine-tune it for, not to out-think closed frontier models across the board - precisely the tradeoff the MoE efficiency bet was designed to make.

NVIDIA frames the open release as strategic rather than generous. CEO Jensen Huang has said free AI should be great for hardware and great for chips ^[6], a philosophy The Decoder's analysis makes explicit: serious production deployment of Nemotron still means buying NVIDIA GPUs, so giving away the weights functions as demand generation for the silicon underneath ^[4]. Reporting puts NVIDIA's open-weight model investment at roughly $26 billion over five years ^[7], and the free framing has a ceiling - while the weights themselves are downloadable at no cost, production-grade support requires an NVIDIA AI Enterprise license priced at $4,500 per GPU per year ^[3], meaning a model that costs nothing to download still funnels revenue back to NVIDIA at scale. Coverage also ties the release's timing to competitive pressure from Chinese open-weight labs such as DeepSeek and Moonshot.ai ^[5], suggesting the open-weight push is as much about contesting that lane as it is about internal hardware economics.

Early enterprise adopters report large, concrete gains: Scale AI's Skydance vendor-credits agent cites roughly 40x lower inference cost and 5x lower latency, Ramp reports 50% higher throughput and 10x lower latency on receipt processing, and AI21's Alice cybersecurity product reports about 40% higher threat-analysis accuracy after fine-tuning Super VL ^[1]. CrowdStrike is building on the family as a named NVIDIA AI design partner for security agents, while Salesforce is separately customizing Nemotron to automate government workflows ^[8][9]. That enterprise optimism doesn't fully match the reception among hands-on hobbyists. Community testing of Bolt 74B-A9B has run skeptical-but-curious: excitement about the 2-bit quantization-aware technique that packs a 74B model into 24GB is tempered by distrust of NVIDIA's model-quality track record, plus a recurring complaint that comparing a 9B-active MoE favorably to 27B-class dense models isn't a fair fight, since MoE trades raw intelligence for speed by design. A widely repeated theory holds that NVIDIA keeps Nemotron deliberately short of frontier-level to avoid cannibalizing its own enterprise GPU sales, noting the release landed the day before RTX Spark's launch. Independent hands-on testing of the smaller Lightning model found throughput results that varied by hardware when compared against a dense 30B model, rather than a uniform speed advantage, and separately found it defaulting to Python despite an explicit instruction to write in C - a small but concrete data point for the specialized-not-frontier read of the family.