How-To Guides

Technical Research Report · Visual Edition

Nemotron 3 vs Gemma 4 31B

Which open model should you deploy?

Benchmarks, GPU footprint, multimodality, license, and a decision tree — read the tradeoff at a glance.

TL;DR — who wins what

Gemma 4 31B takes the academic benchmarks and the clean license; Nemotron 3 owns agentic work, million-token context, and native audio + video. Different halves of "intelligence" — pick by workload, not by leaderboard.

Academic reasoning

Gemma 4 31B

MMLU-Pro 85.2 · GPQA 84.3 · AIME 89.2 — beats even Nemotron Super 120B.

Agentic + long-context

Nemotron 3

SWE-bench 65–70% · Ruler@1M 95% · validated 1M-token window.

Native audio + video

Nano Omni only

The sole open model here doing audio and video, on one GPU.

Cleanest license

Gemma 4 · Apache 2.0

vs NVIDIA's custom (but commercial-OK) Nemotron license.

MMLU-Pro ▸ Gemma GPQA ▸ Gemma AIME ▸ Gemma LiveCodeBench ▸ Gemma SWE-bench ▸ Nemotron Long-context ▸ Nemotron Audio/Video ▸ Nemotron License ▸ Gemma

Scope: the Nemotron 3 family (Nano 31.6B, Super 120B, Ultra 550B, multimodal Nano Omni 30B). "Gemma 4 31B" is real and in production — benchmarked directly; Gemma 3 27B appears only as a faded reference bar.

1. The Nemotron 3 Family

Four open-weight models, Dec 2025 – Jun 2026, all hybrid Mamba-2 + Transformer + MoE (activating ~10% of params per token). From-scratch designs — not pruned from Llama, despite some blogs saying so.

Table 1 — Variants at a glance
Variant	Total	Active	Released	Role
Nano (30B-A3B)	31.6B	3.2B	15 Dec 2025	Efficiency + long-context workhorse. Text-only, single-GPU, ~3.3× Qwen3-30B throughput.
Super (120B-A12B)	120.6B	12.7B	11 Mar 2026	Quality/agentic tier. Text-only. LatentMoE + MTP, 1M context.
Ultra (550B-A55B)	550B	55B	4 Jun 2026	Frontier open reasoning for long-running agents. Text-only. NVFP4 on Blackwell.
Nano Omni (30B-A3B)	~30B	~3B	28 Apr 2026	The multimodal member: vision + audio + video + text, single-GPU.

License at a glance: Nano & Super — NVIDIA Nemotron Open Model License (custom, commercial-OK). Ultra — moving to OpenMDW-1.1 (Linux Foundation). Gemma 4 — Apache 2.0 per Google's official card (Gemma 3 stayed on the custom Gemma Terms). For resale/redistribution, Apache 2.0 is the lowest-friction; verify the exact Nemotron checkpoint license before embedding in a resold product.

2. Benchmarks — Two Different Winners

Gemma 4 31B wins the academic evals; Nemotron 3 wins the agentic and long-context evals. The two charts below show exactly that split.

Chart A — Academic benchmarks (%, higher is better)

Gemma 4 31B Nemotron Super 120B Gemma 3 27B (prior-gen ref)

MMLU-Pro

Gemma 4 31B

85.2

Super 120B

74.1

Gemma 3 27B

52.2

GPQA Diamond

Gemma 4 31B

84.3

Super 120B

79.2*

Gemma 3 27B

24.3

AIME 2026 (math)

Gemma 4 31B

89.2

Super 120B

leads AIME25 — no exact figure published

LiveCodeBench v6 (coding)

Gemma 4 31B

80.0

Super 120B

49.8

MATH

Gemma 4 31B

85.6

Super 120B

78.4

Source: Google official Gemma 4 model card; NVIDIA Nemotron 3 technical reports. * Super GPQA is reasoning-mode-on; sources report a 62–79% range across modes. MATH row: Gemma is MATH-Vision 85.6, Super is MATH 78.4; Gemma 3's nearest published figure is GSM8K 82.6 (different benchmark, omitted). All vendor-self-reported — confirm on your own tasks.

Chart B — Agentic & long-context (%, higher is better)

Nemotron Ultra 550B Nemotron Super 120B Gemma 4 31B

SWE-bench Verified (agentic coding)

Ultra 550B

65–70%

Super 120B

55–60%

Gemma 4 31B

no published source found

Ruler @ 1M tokens (long-context retrieval)

Ultra 550B

Gemma 4 31B

256K max — no validated 1M window

Terminal-Bench 2.0 · IFBench (Ultra)

Terminal-B 2.0

IFBench

τ2-bench (agentic tool-use) — Gemma's turn

Gemma 4 31B

86.4

Nemotron 3

no published source found

Source: NVIDIA Nemotron 3 Ultra technical report & blog; Google Gemma 4 model card. "No published source found" = the vendor did not report a comparable figure (not a zero). The asymmetry is the story: each vendor benchmarks what it optimises for — Nemotron on long-running agents and 1M context, Gemma on single-shot academics and τ2 tool-use.

3. Hardware — What Fits on How Many H100s

MoE models must hold all params in VRAM, not just active ones. The bars show weight memory; the vertical lines show cumulative H100 80GB capacity. Where a bar crosses a line, you need that many GPUs.

Diagram — Weight footprint vs GPU capacity (GB)

BF16 (~2 B/param) FP8 (~1 B/param) NVFP4 / INT4 (~0.5 B/param)

Gemma 4 31B

61 GB · fits 1×

Nemotron Nano

62 GB · fits 1×

Nano Omni

60 GB · fits 1×

Nemotron Super

240 GB · 4×

120 · 2×

60 · 1×

Nemotron Ultra

1,100 GB · >8×H100

550 · 8×H200 / 16×H100

275 · 8×H100 (dequant)

Weights only; add KV cache. Ultra BF16 (1,100 GB, red) overshoots 8×H100 (640 GB) — it needs 8×H200 or 16×H100 across two InfiniBand nodes; FP8 Ultra fits 8×H200 comfortably and 8×H100 only tightly. H100 lacks native FP4 cores, so NVFP4 dequantizes at runtime (native/fastest on Blackwell B200/B300). H100 NVL (dual-GPU, 188 GB) ≈ one NVL pair covers anything up to Super at FP8; Ultra needs ~4 pairs. Throughput (vendor/cloud-reported): Nano ≈3.3× Qwen3-30B on 1×H200; Super ≈1,200 tok/s on 8×H100; Ultra ≈500–900 tok/s FP8 on 8×H200 (5× BF16 at NVFP4 on Blackwell).

4. Multimodality — Capability Matrix

Flagship reasoning models are text-only on both sides. Native audio + video lives only in Nemotron 3 Nano Omni. Gemma 4 does image everywhere, but audio only on its two smallest edge models.

●native ◐partial ○none

Table 2 — Modality support
Modality	Gemma 4 31B	Gemma 4 E2B/E4B	Nemotron Nano/Super/Ultra	Nemotron Nano Omni
Text	●	●	●	●
Image	●	●	○	●
Audio	○	●	○	●
Video	◐	◐	○	●

◐ Video on Gemma = sequences of image frames, not native temporal video. Nano Omni does native vision (dynamic-res images), audio (20 min trained, 5+ hrs inference), and video (Conv3D) in one ~30B model — topping OCRBench v2 (65.8), MMLongBench-Doc (57.5), Video-MME (72.2), VoiceBench (89.4), 5.95 WER ASR. Source: NVIDIA Nano Omni blog/HF; Google Gemma 4 model card.

5. Head-to-Head Scorecard

Each bar leans toward the winner of that dimension. The shape shows the tradeoff instantly: Gemma left, Nemotron right.

◀ advantage: Gemma 4 31Badvantage: Nemotron 3 ▶

Academic reasoning

Math (AIME)

Coding (single-shot)

License openness

Image + text reach

Single-GPU quality

Ecosystem / tooling

Inference efficiency

Fully-open data + recipes

Agentic / tool-use

Long-context (1M)

Native audio + video

Qualitative synthesis of §2–§4 (bar length = strength of advantage, 1–3 steps). Amber = Gemma 4 31B; green/teal = Nemotron 3 (Super/Ultra depending on dimension). Gemma clusters on academics + license + reach; Nemotron clusters on agents, long-context, efficiency, openness, and A/V.

The synthesis: Gemma optimises for reach, license simplicity, and single-GPU/consumer deployment. Nemotron optimises for agentic/long-context performance, MoE efficiency, full data+recipe openness, and native A/V (Nano Omni). Neither is "better" outright — they fit different deployment shapes.

6. Which One? — Decision Tree

Answer top to bottom. The first "Yes" is your model.

Need audio or video understanding?

Yes → Nemotron 3 Nano Omni

▼ no

Long-running agents / tool use / 1M-token context?

Yes → Nemotron 3 Super (→ Ultra)

▼ no

Redistribute / resell the model, or license risk is critical?

Yes → Gemma 4 31B · Apache 2.0

▼ no

Maximise throughput / lowest cost per token on text?

Yes → Nemotron 3 Nano · or Gemma 4 26B MoE

▼ otherwise

Default — single-GPU, on-prem, image + text, best academic scores

Gemma 4 31B

Cost note: MoE VRAM is set by total params (must resident-load all experts), speed by active params. Gemma dense = cheaper to hold, costlier per token; Nemotron MoE = pricier to hold, faster per token. Match the model to whichever resource — VRAM or throughput — is your binding constraint. Cloud H100/H200 ≈ $2–$3.50/GPU-hr (2026).

7. Source Notes & Premise Checks

Table 3 — Where sources needed a note
Item	Resolution
"Gemma 4 31B" exists?	Confirmed real & in production (spring 2026). Benchmarked directly; Gemma 3 27B used only as a reference bar.
Gemma 4 license	Apache 2.0 per Google's official model card — a change from the custom Gemma Terms (Gemma 3 and earlier). Confirmed primary.
Gemma 4 31B benchmarks	MMLU-Pro 85.2, GPQA 84.3, AIME 89.2 (no tools), LiveCodeBench 80.0, MATH-Vision 85.6 — all from the official card. No published IFEval found; τ2-bench 86.4 is the nearest agentic figure.
Nemotron Super lineage	From-scratch hybrid Mamba-MoE (NVIDIA), not pruned from Llama 3.1 405B (an erroneous blog claim).
Super GPQA spread (62 vs 79)	Reasoning-mode-off vs -on; Nemotron exposes a reasoning toggle. Chart uses reasoning-on 79.2.
Nemotron license per checkpoint	Nano/Super custom NVIDIA license; Ultra reportedly OpenMDW-1.1. Verify on the exact HF card.
Ultra on 8×H100	FP8 (~550 GB) is tight on 8×H100 (640 GB); 8×H200 or 16×H100 is the safe config.

Gemma 4 for reach & license; Nemotron 3 for agents, context & audio-video.

Both are strong open models in mid-2026. The decision isn't "which is better" but "which constraint binds." Run a short eval on your real workload before committing — standard practice for any vendor-reported benchmark, applied equally to both.

References

All URLs accessed 9 July 2026. primary = vendor-published (NVIDIA / Google); secondary = third-party, corroborating only.

NVIDIA Nemotron — primary

primary Nemotron 3 family overview — research.nvidia.com/labs/nemotron/Nemotron-3
primary Nemotron 3 Super technical report (PDF) — …/NVIDIA-Nemotron-3-Super-Technical-Report.pdf
primary Nemotron 3 Ultra technical report (PDF) — …/NVIDIA-Nemotron-3-Ultra-Technical-Report.pdf
primary Nemotron 3 Nano technical report (arXiv 2512.20848) — arxiv.org/html/2512.20848v1
primary "Nemotron 3 Ultra Powers Faster Reasoning for Long-Running Agents" — developer.nvidia.com/blog/…nemotron-3-ultra…
primary "Nemotron 3 Nano Omni Powers Multimodal Agent Reasoning" — developer.nvidia.com/blog/…nano-omni…
primary Nano Omni on Hugging Face — huggingface.co/blog/nvidia/nemotron-3-nano-omni-multimodal-intelligence
primary Super 120B-A12B FP8 model card — huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8
primary Nano 30B-A3B FP8 model card — huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8
primary NVIDIA newsroom — Nemotron 3 debut — nvidianews.nvidia.com/news/nvidia-debuts-nemotron-3-family-of-open-models
primary Llama Nemotron Nano VL (8B, 2025) — build.nvidia.com/nvidia/llama-3.1-nemotron-nano-vl-8b-v1/modelcard

Google Gemma — primary

primary Gemma 4 official model card (31B/26B benchmarks, 256K, Apache 2.0) — ai.google.dev/gemma/docs/core/model_card_4
primary Gemma 4 model page — deepmind.google/models/gemma/gemma-4
primary Gemma releases / dates — ai.google.dev/gemma/docs/releases
primary Gemma 4 31B-IT card (NVIDIA build mirror) — build.nvidia.com/google/gemma-4-31b-it/modelcard
primary Gemma 3 27B card (prior-gen ref) — huggingface.co/google/gemma-3-27b-it
primary Gemma Terms of Use (prior-gen license) — ai.google.dev/gemma/terms

Nemotron — secondary / corroborating (deployment & benchmarks)

secondary Spheron — Deploy Nemotron 3 Ultra (H100/H200 sizing) — spheron.network/blog/deploy-nemotron-3-ultra-gpu-cloud
secondary Spheron — Self-Host Nemotron 3 Super — spheron.network/blog/nemotron-3-super-deployment-guide
secondary E2E Networks — Running Super 120B — e2enetworks.com/blog/running-nvidia-nemotron-3-super-120B-on-E2E-networks
secondary DigitalApplied — Super 120B (contains erroneous Llama-lineage claim) — digitalapplied.com/blog/nemotron-3-super-120b-nvidia-open-source-coding-model
secondary DataCamp — Nemotron 3 architecture & benchmarks — datacamp.com/blog/nvidia-nemotron-3
secondary BuildFastWithAI — Ultra review — buildfastwithai.com/blogs/nvidia-nemotron-3-ultra-review-2026
secondary Kaitchup — Nemotron 3 Nano review — kaitchup.substack.com

Gemma — secondary / corroborating (deployment, VRAM, ecosystem)

secondary AurigaIT — Gemma 4 guide (VRAM/quantization) — aurigait.com/blog/gemma-4-features-benchmarks-guide
secondary Labellerr — Gemma 4 technical overview (architecture) — labellerr.com/blog/gemma-4-open-weight-ai-model-overview
secondary Gemma4All — Gemma 4 benchmarks (corroborates official) — gemma4all.com/blog/gemma-4-benchmarks-performance
secondary BenchLM — Gemma 4 31B benchmarks/speed — benchlm.ai/models/gemma-4-31b
secondary MindStudio — Gemma 4 Apache 2.0 license analysis — mindstudio.ai/blog/gemma-4-apache-2-license-commercial-use
secondary WCR.legal — prior Gemma license risks — wcr.legal/google-gemma-license-risks