Nanbeige4.3-33B vs Nemotron 3.5 Lightning 30B A3B: A Teacher Against an Executor

NVIDIA Nemotron 3.5 Lightning 30B A3B runs at 294.9 output tokens per second. Nanbeige4.3-33B has no published speed at all, because nobody serves it. Those two facts tell you more about this matchup than any benchmark row: one model is an engineered throughput product with a rate card and a single-GPU deployment story, and the other is a 33-billion-parameter dense checkpoint that landed on Hugging Face on October 24, 2026 with a model card and no inference provider. They are both "30B class." They are not the same kind of object.

The comparison is worth making because both are being pitched at the same buyer — someone building a long-running agent that needs to execute a lot of small steps without a frontier-model bill. NVIDIA's answer is to shrink the compute per token. Nanbeige's answer is to make a bigger, denser model good enough at agentic work that you only need one. Those are opposite bets, and the evidence for each is very different in kind.

3.6 billion active parameters is the whole argument

Nemotron 3.5 Lightning 30B A3B is a hybrid mixture-of-experts model: 31.6 billion total parameters, 3.6 billion active per token, spread across 52 layers — 23 Mamba-2, 23 MoE, and 6 attention. It routes to 128 experts and activates 6 of them plus a shared expert. That architecture is why it can sustain roughly 295 tokens per second on a managed endpoint while a dense 30B-class model typically lands in the 40–80 range.

Nanbeige4.3-33B is dense. All 33.1 billion parameters fire on every token. There is no routing, no expert sparsity, no Mamba layer. That is a deliberate choice with a clear consequence: each token costs roughly nine times the active compute of Lightning's, and the model needs about 66GB of weights in BF16 — or around 20GB at 4-bit — resident before it generates anything.

Neither design is better in the abstract. Dense models tend to be simpler to fine-tune and more predictable in quality; sparse models tend to be cheaper and faster to serve at the same nominal size. The question is which property your workload is short of.

What each model was built to do

NVIDIA positions Lightning as the execution layer of an always-on agent — the model that takes a plan from a larger system and grinds through tool calls, file edits and browser steps at low cost. It is distilled from Nemotron 3 Ultra, ships with multi-token-prediction speculative decoding baked into pretraining plus separate draft models, and is tuned for harnesses rather than chat. Its published benchmark set reflects that: PinchBench 85.4–86, BrowseComp 36.8–36.97, SWE-bench Verified 51.56–52.8, Terminal-Bench 2.1 around 23.5–24.6.

Nanbeige4.3-33B is positioned as a teacher. The card says so directly: it is the flagship of the 4.3 series, intended to distil into smaller models such as the unreleased Nanbeige4.3-3B. Its benchmark set is correspondingly broad — 20 rows covering agentic coding, agentic cowork, and reasoning — and its recommended settings are those of a model you run yourself: tensor-parallel size 4, XML tool calls, preserve_thinking=true for multi-turn tool use.

One is a component in someone else's system. The other is a source model for a family. That framing explains almost every difference below.

The one benchmark they share cannot be subtracted

Both cards report Terminal-Bench 2.1. Nanbeige claims 80.9. NVIDIA's published range for Lightning is roughly 23.5–24.6. On its face that is a 56-point gap, and it is the number that will get quoted.

It should not be. Terminal-Bench measures a model inside an agent scaffold, and the two vendors do not describe the same scaffold. Nanbeige's card says its Terminal-Bench 2.1 run used Terminus 2 and that its models were evaluated at the highest thinking effort with a shared harness and shared inference parameters. NVIDIA's numbers come from its own internal harness, which the company itself flags as potentially differing from third-party results. A dense 33B model given a favourable loop and a 3.6B-active model given a stock one can differ by tens of points without either model being misrepresented.

The honest reading is directional, not numerical: Nanbeige is claiming substantially stronger agentic terminal performance, and the claim is unverified on both sides. The same caution applies to GPQA Diamond (Nanbeige 89.5, Lightning 74.3–75.6) and to AA-LCR (Nanbeige 84.0, Lightning 52.0–60.3). Those gaps are large enough that some of the difference is probably real capability rather than harness. But "probably" is doing a lot of work in that sentence, and no third party has run the comparison.

What is not ambiguous is independent coverage. Artificial Analysis scores Nemotron 3.5 Lightning at 13 on its Intelligence Index. It has no entry for Nanbeige4.3-33B. One model has a second opinion; the other has a model card. That asymmetry is information, and it cuts against the vendor with the louder table.

1M context on paper, 256K in practice

Artificial Analysis lists Lightning's context window at 1 million tokens. NVIDIA's own documentation is more conservative: 262,144 tokens is the practical ceiling on a single H100. Both numbers can be true — the architecture supports a longer window, the hardware you are likely to run it on does not comfortably hold it.

Nanbeige4.3-33B's 524,288 tokens is a trained window on a dense model, and the memory cost of the KV cache scales with it. The card's vLLM and SGLang recipes use 0.8 GPU memory utilization and tensor-parallel size 4, which tells you the intended deployment is not a single card.

For agent work, the practical difference is less about the maximum and more about what survives a long run. Lightning's hybrid Mamba layers are designed to keep state cheap over long sequences; Nanbeige's dense attention is not. NVIDIA also publishes AA-LCR at 52.0–60.3, which is a long-context reasoning score, and Nanbeige publishes 84.0 on the same benchmark. If you believe both tables, Nanbeige's long-context reasoning is much stronger. If you believe only the independently administered one, Lightning's is the only one you can check.

The arithmetic of hosting a dense 33B

This is where the comparison stops being about quality and becomes about utilization.

Lightning is listed at $0.06 per million input tokens and $0.20 per million output on a managed endpoint. Nanbeige has no rate card because no provider deploys it; the card says that outright. To use it you rent or buy GPUs and follow the four-way tensor-parallel recipe.

Assume, purely for arithmetic, that four H100-class GPUs cost $2 each per hour — $8 per hour for the node. To break even against Lightning's output price, that node has to produce 40 million output tokens per hour, which is about 11,100 tokens per second sustained. A dense 33B model on four GPUs will not do that. Even at a tenth of the GPU cost, the break-even throughput is over a thousand tokens per second sustained, which is a serious serving operation rather than a side experiment.

The conclusion is not that Nanbeige is expensive. It is that self-hosting a dense 33B only beats a $0.20-per-million-token executor at very high, very steady utilization — or when the work cannot leave your infrastructure, in which case the comparison was never about price. At bursty or low volume, the managed sparse model wins on cost by a wide margin before quality enters the discussion.

Where a router fits a two-model system

The architecture NVIDIA is selling — a frontier planner, a cheap executor — is the architecture most agent teams end up with anyway. It is also the shape a routing layer is built for: our routing DSL composes several models into a single call so a planner and an executor can sit behind one endpoint, and automatic failover keeps a struggling executor from turning into a stalled agent.

We do not carry Nemotron 3.5 Lightning 30B A3B or Nanbeige4.3-33B today, so this is not an availability pitch for either. It is a note about where the integration cost lands. If you self-host Nanbeige as the planner and call a hosted model as the executor, the split is worth expressing as routing rules rather than hard-coded in your application — because the model you want in each slot will change before the quarter does.

Latency budget decides more than the leaderboard

• If a single agent step must return in under a second: Lightning. At 295 tokens per second, a 300-token answer lands in about a second; a dense 33B on your own hardware will not, and Nanbeige has no hosted option to benchmark against.

• If the step is a long reasoning pass at the head of a workflow: Nanbeige's claimed margins on GPQA, AA-LCR and the agentic rows argue for it — subject to the harness caveat above.

• If the work is Chinese-language: Nanbeige. Lightning's published languages are English, Spanish, French, German, Italian and Japanese.

• If the work is multilingual European: Lightning. Nanbeige lists English and Chinese only.

• If you have no GPUs and no appetite for them: Lightning, without a contest.

• If data cannot leave your network: Nanbeige, and the cost comparison becomes an infrastructure decision rather than a token-price one.

The open question

Nanbeige4.3-33B is two days old and has no serving story. That is the fact that dates this article. If a provider picks it up — and the card's claim that none does today is the thing to watch — the cost side of this comparison changes completely, because a hosted dense 33B at a competitive per-token price would undercut the entire argument for a smaller sparse executor on quality-sensitive work.

Until then, the defensible position is narrow. Nemotron 3.5 Lightning 30B A3B is the model you can deploy today, at a published price, with an independent score and roughly nine times less compute per token. Nanbeige4.3-33B is the more capable-sounding model on paper, with a 512K native window and self-reported agentic numbers that no one has reproduced. If your agent is latency- and cost-bound, the choice is already made. If it is quality-bound and you own the hardware, the choice is worth testing — and the test has not been run yet.