Nemotron 3.5 Lightning vs gpt-oss-120b: A Cross-Vendor Open Model Face-off

Nemotron 3.5 Lightning is NVIDIA's fast 30B open model, and gpt-oss-120b is OpenAI's larger open model. NVIDIA names gpt-oss-120b as its own comparator, claiming comparable performance at roughly one quarter of the total parameters.

That claim is the heart of this matchup. A 30B model that matches a 120B model would change the math on hardware, speed, and cost to run.

The short answer: Lightning is the efficiency play, and gpt-oss-120b is the pick when you want the largest total capacity or already run OpenAI's open model. Both are open weights with no per-token fee when self-hosted.

This guide compares the two on parameters, architecture, speed, benchmarks, license, and self-hosting cost. Both are open weights, so the real question is which one delivers more per dollar of hardware.

Nemotron 3.5 Lightning vs. gpt-oss-120b: Side-by-Side

The Core Claim: Same Class, One Quarter the Parameters

NVIDIA's headline claim is that Nemotron 3.5 Lightning matches gpt-oss-120b at roughly one quarter of the total parameters. Lightning has 30B total parameters against gpt-oss-120b's 120B, which lines up with that one-quarter figure.

If the performance really is comparable, the smaller model wins on efficiency. Fewer total parameters mean smaller memory footprint and, with only 3B active, cheaper inference.

Read the claim as NVIDIA's published position, not a neutral test. It is a strong marketing point, and it is the exact kind of assertion you should confirm on your own tasks.

Comparable does not mean identical. A smaller model can match an average score yet differ on specific tasks, so treat the one-quarter figure as a reason to test, not a reason to skip testing.

Architecture and Active Parameters

Lightning uses a hybrid design: Mamba-2 state-space layers, MoE layers, and a few Attention layers. It activates 3B of its 30B parameters per token.

gpt-oss-120b is an open Mixture-of-Experts model from OpenAI with a far larger total parameter count. For its exact active count and layer layout, check its official model card rather than assume.

Active parameters drive inference cost more than total size. A model with a low active count can run fast even when its total is large, so compare active counts, not just the headline totals.

Total size still sets the memory floor. All 120B of gpt-oss-120b's parameters must load into GPU memory to serve, even if only some activate per token. Lightning's 30B total is what lets it fit on smaller hardware, which is the practical meaning of NVIDIA's one-quarter claim.

Speed and Benchmarks

Lightning is tuned for output speed, with NVIDIA claiming up to 4x throughput versus similar-sized models. That speed is part of why NVIDIA frames it as an efficient alternative to a 120B model.

For quality, NVIDIA publishes Lightning BF16 scores of SWE-bench Verified 51.56, GPQA Diamond 75.44, MMLU Pro 81.94, and PinchBench 85.37. It reports these as comparable to gpt-oss-120b while using far fewer total parameters.

Do not take cross-model scores at face value. Vendors run benchmarks under their own settings, so verify both models on the same tasks and hardware before you decide.

Speed and quality trade off in practice. A model that is faster but slightly less accurate can still win on cost per useful task, or lose it, depending on how often you must retry. Measure both together on your workload.

License and Cost to Self-host

Both models are open weights, so self-hosting has no per-token fee. You pay for GPUs, power, and operations, not for each token you generate.

Here Lightning's smaller size is a real cost lever. A 30B model with 3B active fits on more modest hardware than a 120B model, which lowers the barrier to run it on RTX or DGX systems, or through vLLM.

Across the model launches we track, total parameter count is the quiet driver of self-hosting cost. If two models perform alike, the smaller one is usually cheaper to serve at scale, which is the case NVIDIA is making here. Check gpt-oss-120b's own license terms before you deploy it.

→ Compare per-token pricing for every AI model

Checkpoints, Training, and Serving

Nemotron 3.5 Lightning ships in NVFP4 4-bit and BF16 checkpoints, both on Hugging Face. The 4-bit build cuts memory needs so the model fits on smaller GPUs, while BF16 keeps higher precision.

NVIDIA pre-trained Lightning on more than 20 trillion tokens. You can run it as a hosted API on build.nvidia.com, locally on RTX or DGX hardware, or through open serving stacks like vLLM.

gpt-oss-120b is also open weight and widely supported across serving tools. Its larger total size means it needs more GPU memory to serve at full precision, which raises the hardware bar.

NVIDIA shipped NeMo Switchyard alongside Lightning. It is a separate model router, not the model itself, and it can direct requests across models in a mixed stack that includes both open models.

What a Switch Actually Involves

Moving from gpt-oss-120b to Lightning is mostly a serving and prompt exercise, not a rewrite. Both are open weights you host yourself, so your application code and tool interfaces can stay largely the same.

The real work is re-tuning prompts and checking quality. A smaller model can behave differently on edge cases, so you need side-by-side tests on your own tasks before you cut over.

The payoff is hardware. If Lightning matches quality at a quarter of the parameters, you can serve the same workload on fewer or smaller GPUs, which lowers ongoing cost.

Keep a rollback path during the switch. Run both models in parallel on a slice of traffic, compare results, and only retire gpt-oss-120b once Lightning clears your accuracy bar.

Which One Fits Your Stack

Choose Lightning when efficiency matters most. Fast agent loops, tight hardware budgets, and long-context runs all favor a small-active 30B model.

Choose gpt-oss-120b when you are already standardized on OpenAI's open model, or when you want the largest total capacity and have the hardware to serve it. Some teams value staying within one vendor's tooling.

There is a headroom argument for the larger model. If your tasks include hard, open-ended reasoning where extra capacity helps, a 120B model gives you more room to grow. Match that against the higher serving cost, since capacity you never use is capacity you still pay to host.

For most agent workloads, the deciding factor is cost per useful task. If Lightning matches quality at a quarter of the parameters, it is the cheaper engine, but only your own test proves the match holds.

There is also a hardware-availability angle. A 30B model with 3B active runs on a wider range of GPUs, so teams without top-end clusters can serve it more easily. gpt-oss-120b needs more memory, which can limit where you deploy it and how many replicas you run.

The Verdict

Pick Nemotron 3.5 Lightning when efficiency is the goal. It carries one quarter of gpt-oss-120b's total parameters, activates only 3B, and NVIDIA claims comparable performance with up to 4x output speed. That combination makes it cheaper to self-host on modest hardware and well suited to high-volume agent loops.

Pick gpt-oss-120b when you are already committed to OpenAI's open model, want the largest total capacity, and have the GPUs to serve a 120B model. Some teams prefer to keep one vendor's ecosystem even at a higher hardware cost.

Because both are open weights with no per-token fee when self-hosted, the choice comes down to cost per useful task. Run both on your own workload, at the same precision and hardware, and confirm NVIDIA's comparable-performance claim before you commit. Verify all benchmark numbers on each official model card.