An open 120B model that activates 12B per token
NVIDIA's blog says it released Nemotron 3.5 Super VL today, an open 120-billion-parameter model, and that it delivers the fastest task completion and the highest accuracy for specialized always-on agents. The blog time is 15:00 UTC on October 15. The technical blog, 18:01 UTC the same day, describes a mixture of experts: 120B parameters, 12B active per token, combining reasoning, instruction following, tool use, and text, image, document, and video understanding in one model.
The speed and accuracy superlatives are the blog's wording. The technical blog narrows the throughput claim: on a single GPU, up to 5 times the throughput of other open models in its size category on agentic tasks. That sentence does not name the other model. The blog also calls it the fastest open agentic model in its class, 5 times faster than the next best. The next best is not named there either.
[1][2]
The 24GB Bolt is out; the 4GB one is dated October 23
Two edge models were announced with it. Nemotron 3.5 Bolt 74B-A9B 24GB is downloadable now. A native 2-bit quantization fits it in 24GB of memory so it can run alongside other applications. The blog says that, relative to models with a similar memory footprint, it is the number-one open model in its class on PinchBench and the number-two open model on BrowseComp. The comparison list is not in the post. Bolt E4B 4GB is written as coming on October 23. It is not a downloadable release yet.
The blog says Super VL and Bolt 74B-A9B are available now, in the same family as Lightning, released in August, and Ultra, which handles the hardest reasoning.
[1]The chart scores are against NVIDIA's own models
The technical blog says that on CharXiv, a chart-reasoning benchmark, the model scored 80.8 percent, against 63 percent for Nemotron 3 Nano Omni, and 84 percent with Python code interpretation. On Hermes Bench and LangChain behavioral evaluations, Super VL scores 63.44 percent and 87.05 percent, against 56.33 percent and 84.3 percent for Nemotron 3 Ultra. The technical blog says Super VL scores higher than Ultra on those two. The comparison is with NVIDIA's own models, not an external overall board.
The blog says pretraining and continued pretraining used about 26 trillion tokens, and post-training used roughly 3.8 million samples, plus reinforcement learning across dozens of environments. Open datasets are on Hugging Face, covering tool use, finance, instruction following, long context, fairness, and vision-language understanding.
[1][2]Some of the customer multiples are still in the future tense
Scale AI post-trained the model for vendor-credit processing at Skydance. The technical blog says inference cost for this workload will decrease by 40 times and latency by 5 times. That is prospective, not a closed bill. Ramp is turning receipt images into JSON with merchant, date, and payment fields, and expects up to 50 percent higher throughput and 10 times lower latency. AI21 and Alice post-trained it for threat intelligence and vulnerability analysis, improving accuracy by 41 percent. The 41 percent has no sample size and no baseline score. An automatic summary at the top of the technical page folds the three into one line about 40 times cost, 10 times latency, and 40 percent accuracy. The body does not. They stay separate here.
On DGX Spark, the technical blog's recommended LiLiCorr drafter is up to 2.3 times faster at token generation than decoding without speculation. On 16 B200 GPUs, at 50 tokens per second per user, KV-aware routing with session affinity and KV-cache offload serves up to 8 times as many users and cuts first-token latency by 85 percent. Those are NVIDIA's own serving figures.
[1][2]