Granite Switch 4.1 30B vs Nemotron 3.5 Lightning 30B A3B: What "Switch" Means in Each Name

Both of these models are 30B-ish, both have a word in the name that sounds like routing, and they are doing completely different things with it. Granite Switch 4.1 30B — weights on the Hub since 14 October 2026 — switches adapters: one dense base model, two activated-LoRA attachments, every token running the full network. Nemotron 3.5 Lightning 30B A3B, which NVIDIA published on 11 August 2026, switches experts: a hybrid Mamba-2 and MoE stack where only about 3 billion of the 30 billion parameters fire per token. Nobody has benchmarked the two against each other and this piece is not going to pretend otherwise. The genuinely useful comparison is what each design costs you per token and what it hands back.

Getting the distinction right is not pedantry, because the names invite exactly the wrong inference. "Switch" in Granite Switch refers to a lightweight switch layer that reads control tokens in the input and tells each decoder layer which LoRA weights to apply over that span of tokens. It is not mixture-of-experts routing and it does not reduce the compute per token by a single FLOP. "A3B" in NVIDIA's name is the opposite: it is an explicit claim about how little of the model runs at inference time.

The scoreboard, dimension by dimension

• Parameters — Granite Switch 4.1 30B composes to 29,989,814,272 over a 28,865,728,512 base (+3.89%); Nemotron 3.5 Lightning 30B A3B is 30B total with roughly 3B active per token.

• Architecture — dense decoder-only transformer plus two activated-LoRA adapters versus an interleaved Mamba-2, MoE and attention hybrid.

• Compute per token — full dense 30B forward pass versus roughly a tenth of that in active parameters, which is the entire point of the A3B suffix.

• Context — 131,072 tokens on the Switch checkpoint; up to 1M on Nemotron Lightning, with NVIDIA documenting 256K as the practical figure on a single H100.

• Reasoning mode — none in the Switch chat template; a configurable enable_thinking toggle in Nemotron's.

• Audio input — an embedded 470M speech encoder on the Switch model; none on the NVIDIA side.

• Licence — Apache-2.0 versus OpenMDW-1.1. These are not the same document and the downstream terms differ; read the one you are actually shipping under.

• Hardware — the Switch card lists Transformers and vLLM within a narrow version window; NVIDIA documents Nemotron Lightning on a single H100 or A100 80GB, with an NVFP4 build intended for deployment.

What the Granite Switch checkpoint is actually buying you

The composition is the product. Granite Switch 4.1 30B freezes Granite 4.1 30B and attaches two adapters — a rank-16 guardian-core safety classifier on the attention projections and a rank-64 decision adapter across all seven projections — plus the speech encoder. Adapter selection happens through the chat template when you pass an adapter_name, so it is a request parameter, not a second service.

What that buys is capability density: a safety verdict, a calibrated option distribution and a transcript out of one process, for 1.12 billion extra parameters on top of a base you were going to load anyway. Six months ago the same project shipped a preview checkpoint with twelve adapters at an 11.68% surcharge; the finished version cut that to two at 3.89%, which is IBM trading coverage for legibility. The other ten adapters still exist and the composer will build you a checkpoint with any subset.

The costs sit exactly where you would expect. Nothing about the adapter machinery makes the base model faster, because the base model is dense and untouched — every token still runs the whole network. The speech path is English-only, lowercase, unpunctuated and, in IBM's own words, not jointly trained with the language model, so a mistranscription silently changes the question the model answers. And the checkpoint ships in one precision with no official quantized build, against an NVIDIA model whose deploy-recommended NVFP4 variant has been on the Hub since early August.

What the Nemotron numbers are actually measuring

NVIDIA's benchmark table for Nemotron 3.5 Lightning 30B A3B is fuller than IBM's is for the Switch model, and it is worth reading the footnote attached to it. The figures — MMLU-Pro 81.94, GPQA Diamond 75.44, SWE-bench Verified 51.56, Terminal-Bench 2.1 24.58, PinchBench 85.37, AA-LCR 52.00 on long context — are described by NVIDIA as measured under its own consistent harness, with the caveat that they may differ from other vendors' self-reported numbers. That is a more honest framing than most cards manage, and it still leaves them vendor-reported. NVIDIA published the recipes in NeMo Gym so the runs are reproducible by someone else, which is not the same as someone else having reproduced them.

The more interesting claim is the one about speed rather than accuracy. A 30B model with 3B active parameters is a bet that most agentic work does not need the whole network to fire on every token, and NVIDIA shipped speculative decoding alongside it — multi-token prediction heads plus the DFlash and DSpark drafters — to push throughput further. It also positions Lightning as an execution layer beneath a larger planner rather than as the model you ask hard questions. That positioning is NVIDIA's, and it is the right lens for the benchmark table too: Token-level efficiency is what shows up in a bill.

The trap in putting these two scoreboards together

There is one number published for the Switch model's headline capability — the Decision Index, edition v0.3, public fast subset of 1,068 rows, Overall 50.1, with Knowledge 32.4, Language 58.6, Retrieval 49.8, Tools 76.5 and Arts 26.4 — and there is a full benchmark table for Nemotron Lightning. Placing 50.1 beside 81.94 or 51.56 would be a category error of the purest kind.

The Decision Index scores how well the decision adapter ranks a fixed set of labelled options. It is not a measure of the base model's chat or reasoning quality, and the base model is unmodified by the adapter. Nemotron's table scores the whole model on knowledge, reasoning, coding and long context. Different axes, different harnesses, different authors, and no shared benchmark between them. Both are vendor-reported. Neither is comparable to the other, and the only honest thing to do with the pair is refuse to subtract them.

What you can legitimately do is read the shape of each. The Decision Index's 76.5 on Tools against 26.4 on Arts says the adapter is worth pointing at tool-selection decisions and should not be trusted on open-domain classification without your own evaluation. Nemotron's 51.56 on SWE-bench Verified against a same-table best of 70.12 tells you it is not the strongest coding agent in its own comparison set — NVIDIA put a competitor's higher number in its own table, which is at least a signal worth respecting.

Which one, for what

If your problem is a long agent loop running thousands of short tool decisions, the parameter economics point one way: a 3B-active MoE on a single GPU will do that work at a fraction of the cost per token of a dense 30B, and the Switch model has no mechanism to close that gap, because its "switch" never reduces the compute. Add the 1M context and the reasoning toggle and NVIDIA's model is simply built for that workload.

If your problem is one process that must answer a user, emit a calibrated routing decision and transcribe what the user said, the arithmetic inverts. Nemotron 3.5 Lightning has no decision head and no audio path at all, and prompting a chat model to please reply with one letter is not a calibrated distribution. Granite Switch 4.1 30B hands you all three out of one checkpoint for under 4% more weights than its base, and that is a capability you cannot buy by picking a better model.

The honest third case is the most common one: you are not choosing between these at all. Both are self-host artifacts, both are Apache-2.0 or OpenMDW-1.1 rather than a hosted endpoint, and neither has been independently evaluated. Granite Switch 4.1 30B is two days old with a single published figure; Nemotron 3.5 Lightning is two months old with a vendor table and published recipes. That is not enough to plan a migration on.

Getting a real answer without a migration

Neither model is in the OrcaRouter catalogue today, so both mean your own hardware and the vendor's own release channels — IBM's for the Switch checkpoint, NVIDIA's for Lightning. Saying otherwise would be inventing an endpoint.

Where the platform earns its place is that a candidate should be cheap to test. One key covers 200-plus models with automatic failover and passes provider list price through with 0% markup, so when a hosted route to either model appears, putting it beside the model already in production is a configuration change rather than a procurement cycle. Faced with two 30B open-weight models that both look reasonable on paper and neither of which anyone outside the respective vendor has measured, the decision that matters is not which card you believe. It is how cheaply you can find out.

Bottom line

Granite Switch 4.1 30B and Nemotron 3.5 Lightning 30B A3B share a parameter count and almost nothing else. One is a dense base with switchable adapters that gives you three jobs per process at full dense compute per token; the other is a sparse hybrid built to make each token cost as little as possible and to run a long loop on a single GPU. Choose on the workload's shape, not on a benchmark table — and if you take one thing from this, take the correction in the names: adapter switching buys you capability, expert sparsity buys you throughput, and no model here buys you both.