Three billion of thirty billion is ten percent [16], and against Qwen3.8-27B's dense 27 billion it is nine times fewer parameters touched per token [17]. That is a bandwidth claim rather than a capacity claim. Memory still has to hold every expert, because the next token may route to any of them; only the per-token read shrinks. So the first thing to check on a given module is whether the quantized checkpoint plus your KV cache fits at your context length. NVFP4 is the lever the post points at, described as reducing both the work and the memory required for model operations [9].
NVIDIA maps the two architectures onto different workloads: Nemotron 3.5 Lightning for response-heavy flows where faster token generation shortens the overall process, Qwen3.8-27B for tasks with fewer, harder decisions that allow more time per response [7]. Read that as a statement about your token mix. An agent emitting many short tool calls and short verifications lives in decode, where activated parameters set the pace. One that writes a long plan once an hour does not.
The performance figure is where transfer gets hard. Figure 2 compares both models in BF16, in NVFP4, and in NVFP4 with the fastest speculative decoding configuration NVIDIA tested for each model [11]. Those final bars come from two separately tuned configurations, not one shared setup. Speculative decoding pays when drafted tokens are accepted at verification [10], and acceptance is a property of your output distribution rather than of the silicon: schema-shaped tool calls draft well, free-form reasoning drafts worse. For a published speedup to survive contact with your service you need the same draft model, similar acceptance on your prompts, and the same concurrency, since a draft length tuned for one request at a time is the wrong one under a batch. The plainest line in the post says roughly this: benchmark both models on the decisions, tools, and response patterns your application requires before selecting one [12].
The routing decision is narrower than the framing suggests. The reference loop has an agent monitor a system from live sensor data and device logs, take approved corrective actions, verify results against predefined tests, and escalate to an expert only when needed [13]. Escalation is a network call. NVIDIA's careful phrasing is that deploying locally through vLLM or llama.cpp means the reasoning loop "does not depend entirely on the data center" [14]. Two words carrying an entire hybrid architecture. If your round trips exist because data has to stay on device [1], what changes is which calls leave, not that none do.
The capability evidence is an aggregate index. Figure 1 plots the Artificial Analysis Intelligence Index against model size and release date, with 2026 open models reaching scores similar to leading 2025 models at far fewer parameters [3], and part of the mechanism named for the smaller model is distillation from Nemotron 3 Ultra [4]. That index does not measure tool-calling accuracy after quantization on your own schemas. That measurement is left to the adopter, and it is the one that decides whether the data center endpoint gets retired or just demoted.
