Key takeaway: Nemotron 3.5 Lightning is built to do the dull execution steps in an agent — tool calls, checks, delegation — not to be the cleverest model in the room. That's the tell.
Why it matters: The economics of agents break when you send every step to a huge reasoning model. Specific and cheap beats smart and slow, and that reshapes how marketing teams should build.
What happened
Nvidia has released Nemotron 3.5 Lightning, a 30-billion-parameter model with fully open weights. It arrived, per Business Standard's report on the Nemotron launch, a day after Meta shipped its own 30B open model, Muse Glimmer.
The interesting part isn't the size. According to Business Standard, Lightning is designed for the repetitive execution tasks agents do all day — making tool calls, validating results, and handing work to subagents. Nvidia's point: running a large reasoning model at every one of those steps is costly and slow. It uses a mixture-of-experts design with only about 3 billion parameters active per token.
The backdrop is a genuine shift. Data from Vercel's production AI gateway found open-weight models processed 29% of all tokens in June 2026, up from roughly one-ninth two months earlier.
Source: Computing / Vercel AI Gateway Production Index, 2026
Everyone will say Nvidia is buying ecosystem lock-in with free models
The consensus take writes itself, and it's fair. Nvidia sells the GPUs; giving away capable models that run best on those GPUs deepens developer adoption and keeps the hardware flywheel spinning. Free models, more demand for chips, more moat. It's clever strategy dressed as generosity, and the "picks and shovels" firm is now shaping what runs on the shovels.
That reading is true. It's just not the most useful thing to notice if you actually build with this stuff.
Our take: the boring agent won, and this release is the receipt
We think the real story is architectural, not commercial. For two years the pitch was one enormous model that could do anything. Nvidia has just shipped a small, fast, cheap model whose whole job is to grind through the unglamorous middle of an agent's work. That's not a compromise. That's the direction of travel.
In our experience building agents, the thing that breaks in production is never the clever bit. It's the hundredth tool call, the retry loop, the validation step that runs a thousand times an hour. Sending each of those to a top-tier reasoning model is like hiring a barrister to file your post. The costs are brutal — some analyses put agentic workflows at 5–30x the token spend of a single answer, which is exactly the trap Nvidia is pricing around.
So the winning pattern isn't "smartest model, everywhere". It's a narrow, reliable model doing the repetitive execution, with a bigger model reserved for the handful of steps that genuinely need judgement. Trust comes from scope, not IQ. An agent you can trust is one that does a small, well-defined job the same way every time — and a 3B-active model doing tool calls is a very boring, very trustworthy little worker.
This is why we build the way we do. When we design AI agents for marketing, we don't reach for one giant model and hope. We give each step the smallest capable model that reliably does the job, and we let the orchestration decide what goes where. Nvidia even shipped a routing layer, NeMo Switchyard, to do precisely this — an admission that the model is not the product; the routing between models is.
The hype cycle wants you to buy the biggest brain. The engineering reality, which Nvidia has now stamped its logo on, is that the biggest brain is the wrong default for most of what an agent does. Narrow beats broad. Specific beats smart. That's the whole game.
What this means for marketing teams
- Audit where your agents actually spend tokens. If a large model is handling routine tool calls, you're likely overpaying by a multiple — the 5–30x figure is a live risk, not a scare story.
- Split your workflow by step, not by tool. Route the repetitive execution to a small open-weight AI model and reserve the expensive reasoning for the few steps that need it.
- Treat open weights as an option, not an ideology. With open models now near a third of production traffic, self-hosting for privacy or cost is a real lever — but only where it earns its keep.
- Measure reliability per step over a two-week window before you scale. A boring agent that's right 99% of the time beats a clever one that's right 90%.
- If you'd rather not stitch the routing yourself, our pricing and plans lay out how we handle it — no giant-model tax by default.
Frequently asked questions
What is Nvidia Nemotron 3.5 Lightning?
Nemotron 3.5 Lightning is Nvidia's 30-billion-parameter open-weight AI model, released in August 2026. It's built for the high-volume, repetitive execution steps in AI agents, such as tool calls, validation and delegating work to subagents.
Why release a small model instead of a bigger, smarter one?
Because running a large reasoning model at every step of an agent is slow and costly. A smaller model handles routine execution cheaply, reserving expensive reasoning for the few steps that genuinely require judgement.
Are open-weight AI models actually being used in production?
Yes. Data from Vercel's production AI gateway found open-weight models processed 29% of all tokens in June 2026, up from roughly one-ninth two months earlier — a sharp rise in real deployment.
