model
Nvidia's open Nemotron 3.5 Lightning is a 30B-parameter mixture-of-experts model (3B active) that runs on a single consumer GPU.
No evidence-backed relationships are recorded.
Red Hat clocks the same 20-call agent task at roughly 45 seconds on a slow backend and about 13 on a fast one. Model choice for agents is turning into a per-call latency budget, with capability as one input.
Reality
NVIDIA's dense-versus-MoE explainer uses Nemotron 3.5 Lightning to walk through per-layer routing. Its own text puts the throughput number and the memory number on two different parameter counts.
Reality
Gartner puts this year's AI spending at $2.59 trillion, but the software and services layer the labs actually compete for is barely $1 trillion of it, which is what the compressed release calendar is defending.
Reality
Nemotron 3.5 Lightning touches a tenth of its weights per token, which is what makes an agent loop plausible next to the sensors, but memory still has to hold all thirty billion and NVIDIA's own loop still escalates off-device.
Reality
The framework's inputs are token lengths, concurrency, a latency percentile and the share of prompt tokens already sitting in the KV cache, and each of those changes arithmetic that a per-GPU throughput rating leaves untouched.
Reality
Portable Computer runs Perplexity's agent on 27B models on your own hardware. Paid tiers only, Linux first, Windows in September, and a 24GB VRAM floor most desktops miss.
Reality
Portable Computer keeps agent work on hardware you own and asks permission before any single step escalates. The entry fee is 24GB of VRAM and a paid subscription.
Reality
NVIDIA's Nemotron 3.5 Lightning is now deployable from SageMaker JumpStart and fits on one GPU. The argument underneath it is about billing, not benchmarks.
Reality
Grok 4.6, Qwen3.8-Max and DeepSeek V4-Pro shipped inside about 24 hours, and two of the three came with downloadable weights. The benchmarks existed to justify a cheaper invoice.
Perspective Coverage
4 publishersReality
A verify-on-read experiment rerun across 14 live models on a fingerprinted 50-fact set found false-accept rates up to 0.38, and run-to-run noise wide enough to swallow a prompt fix.
Reality